[{"_1":2,"_3":-5,"_4":-5},"loaderData",{"_5":6,"_7":8},"actionData","errors","root",{"_1981":1982,"_1983":-5,"_9":1984,"_1985":1986},"routes/library.video.$slug",{"_9":10,"_11":12,"_13":-5,"_14":15},"data",{"_16":17},"preview",false,"pageNumber","slug","persistent-memory-and-what-breaks-in-production","routeData",{"_18":19,"_20":21,"_22":23,"_24":25,"_26":27,"_28":29,"_30":31,"_32":33,"_34":-5,"_35":36,"_37":-5,"_38":39,"_40":41,"_42":43,"_44":45,"_46":-5,"_47":48,"_49":50,"_51":52,"_53":54,"_55":56,"_57":-5,"_14":58,"_59":-5,"_60":61,"_62":63,"_64":65,"_66":67,"_68":69,"_27":70,"_71":72},"_createdAt","2026-10-23T21:54:05Z","_id","39c370db-08bf-4de5-b2c2-7b579fe65e59","_rev","C6cJy7Ex0HWbukQjnuFRXX","_system",{"_1645":1979},"_type","video","_updatedAt","2026-10-27T02:12:07Z","access","Public","afterEvent",true,"agenda","collections",[],"cta","endDate","2026-09-25T02:00:00.000Z","excerpt",[1972],"filmedOn","2026-09-24","library",{"_1870":1871,"_1872":1873},"modules","moreFromLibrary",[1862,1863,1864],"outlines",[1854],"posterFrame",{"_26":1851,"_77":1852},"publishDate","2026-10-26T15:20:00.000Z","skill",{"_18":1649,"_20":1650,"_22":1651,"_24":1652,"_26":55,"_28":1653,"_1654":1655,"_1543":1656,"_1657":1658,"_1659":1660,"_1661":1662,"_1663":1664,"_14":1665,"_1666":1667,"_59":1668,"_66":1669},"slideDeck",{"_26":14,"_1548":15},"speakers","speakersWithOrg",[1590],"startDate","2026-09-24T16:30:00.000Z","tags",[1532,1533,1534,1535,1536,1537,1538],"title","Persistent Memory + What Breaks in Production","transcript",[153,154,155,156,157,158,159,160,161,162,163,164,165,166,167,168,169,170,171,172,173,174,175,176,177,178,179,180,181,182,183,184,185,186,187,188,189,190,191,192,193,194,195,196,197,198,199,200,201,202],{"_26":76,"_77":78},"seo",{"_73":74,"_75":52},"OGdescription","Sarah Wooders, CTO of Letta and co-author of MemGPT, on how her team works with long-running agents: persistent memory, continual learning in token space, what breaks when context rots and memory gets polluted, and how to measure what an agent actually learned.","OGimage","mux.video","asset",{"_18":79,"_20":80,"_22":81,"_26":82,"_28":83,"_84":85,"_9":86,"_87":88,"_89":90,"_91":92,"_93":94},"2026-10-23T22:03:55Z","80225ef5-5093-4a78-892b-a985c9d2d7ec","E6RGkrOtRyf4APQmHaRL9Z","mux.videoAsset","2026-10-23T22:07:52Z","assetId","6HM8WUbl8XYTM2vYlEUiYaBl9PTcSQ15",{"_95":96,"_97":98,"_99":100,"_101":102,"_103":85,"_104":105,"_106":107,"_108":109,"_110":111,"_112":113,"_114":107,"_115":116,"_117":80,"_118":119,"_120":121,"_122":109,"_89":90,"_123":124,"_125":92,"_126":127},"playbackId","WUC7YYclMdxXDY02rzsYF5AyLHBqXYEHf","status","ready","uploadId","kTpaFq7iGGYqqjEeGJ1TIP24bLIAd4Sl","url","https://stream.mux.com/WUC7YYclMdxXDY02rzsYF5AyLHBqXYEHf","aspect_ratio","16:9","created_at","1792793073","duration",751.461333,"encoding_tier","smart","id","ingest_type","on_demand_direct_upload","master_access","none","max_resolution_tier","1080p","max_stored_frame_rate",29.97,"max_stored_resolution","UHD","mp4_support","non_standard_input_reasons",{"_151":152},"passthrough","playback_ids",[148],"progress",{"_120":145,"_146":147},"resolution_tier","tracks",[128,129],"upload_id","video_quality","plus",{"_99":130,"_103":138,"_139":140,"_141":142,"_143":33,"_137":144},{"_99":130,"_103":131,"_132":111,"_133":134,"_135":136,"_137":27},751.417333,"qLiRmpND014x2nK0067aaSoLsODXwXQPqoemDQP6sPiSg","max_frame_rate","max_height",1152,"max_width",2048,"type","VBzqXd524VWrJ00MwgCHzjFwLPmlPHvpz","max_channel_layout","stereo","max_channels",2,"primary","audio",100,"state","completed",{"_103":88,"_149":150},"policy","public","video_resolution","3840x2160",{"_203":1512,"_26":205,"_206":1513,"_208":1514,"_210":211},{"_203":1482,"_26":205,"_206":1483,"_208":1484,"_210":211},{"_203":1475,"_26":205,"_206":1476,"_208":1477,"_210":509},{"_203":1445,"_26":205,"_206":1446,"_208":1447,"_210":211},{"_203":1415,"_26":205,"_206":1416,"_208":1417,"_210":211},{"_203":1385,"_26":205,"_206":1386,"_208":1387,"_210":378},{"_203":1355,"_26":205,"_206":1356,"_208":1357,"_210":211},{"_203":1315,"_26":205,"_206":1316,"_208":1317,"_210":211},{"_203":1308,"_26":205,"_206":1309,"_208":1310,"_210":509},{"_203":1278,"_26":205,"_206":1279,"_208":1280,"_210":211},{"_203":1258,"_26":205,"_206":1259,"_208":1260,"_210":211},{"_203":1238,"_26":205,"_206":1239,"_208":1240,"_210":211},{"_203":1208,"_26":205,"_206":1209,"_208":1210,"_210":211},{"_203":1178,"_26":205,"_206":1179,"_208":1180,"_210":211},{"_203":1158,"_26":205,"_206":1159,"_208":1160,"_210":211},{"_203":1128,"_26":205,"_206":1129,"_208":1130,"_210":211},{"_203":1108,"_26":205,"_206":1109,"_208":1110,"_210":211},{"_203":1088,"_26":205,"_206":1089,"_208":1090,"_210":211},{"_203":1081,"_26":205,"_206":1082,"_208":1083,"_210":509},{"_203":1061,"_26":205,"_206":1062,"_208":1063,"_210":211},{"_203":1041,"_26":205,"_206":1042,"_208":1043,"_210":378},{"_203":1011,"_26":205,"_206":1012,"_208":1013,"_210":211},{"_203":971,"_26":205,"_206":972,"_208":973,"_210":211},{"_203":941,"_26":205,"_206":942,"_208":943,"_210":211},{"_203":911,"_26":205,"_206":912,"_208":913,"_210":211},{"_203":904,"_26":205,"_206":905,"_208":906,"_210":509},{"_203":884,"_26":205,"_206":885,"_208":886,"_210":211},{"_203":864,"_26":205,"_206":865,"_208":866,"_210":211},{"_203":834,"_26":205,"_206":835,"_208":836,"_210":211},{"_203":814,"_26":205,"_206":815,"_208":816,"_210":378},{"_203":784,"_26":205,"_206":785,"_208":786,"_210":211},{"_203":754,"_26":205,"_206":755,"_208":756,"_210":211},{"_203":704,"_26":205,"_206":705,"_208":706,"_210":211},{"_203":684,"_26":205,"_206":685,"_208":686,"_210":211},{"_203":644,"_26":205,"_206":645,"_208":646,"_210":211},{"_203":614,"_26":205,"_206":615,"_208":616,"_210":211},{"_203":574,"_26":205,"_206":575,"_208":576,"_210":211},{"_203":534,"_26":205,"_206":535,"_208":536,"_210":211},{"_203":514,"_26":205,"_206":515,"_208":516,"_210":211},{"_203":506,"_26":205,"_206":507,"_208":508,"_210":509},{"_203":476,"_26":205,"_206":477,"_208":478,"_210":211},{"_203":456,"_26":205,"_206":457,"_208":458,"_210":211},{"_203":436,"_26":205,"_206":437,"_208":438,"_210":211},{"_203":396,"_26":205,"_206":397,"_208":398,"_210":211},{"_203":375,"_26":205,"_206":376,"_208":377,"_210":378},{"_203":355,"_26":205,"_206":356,"_208":357,"_210":211},{"_203":325,"_26":205,"_206":326,"_208":327,"_210":211},{"_203":295,"_26":205,"_206":296,"_208":297,"_210":211},{"_203":255,"_26":205,"_206":256,"_208":257,"_210":211},{"_203":204,"_26":205,"_206":207,"_208":209,"_210":211},"_key","a94643d653a0","block","children",[226,227,228,229,230,231,232],"markDefs",[212,213,214,215],"style","normal",{"_203":224,"_26":217,"_218":225},{"_203":222,"_26":217,"_218":223},{"_203":220,"_26":217,"_218":221},{"_203":216,"_26":217,"_218":219},"364967e","timestamp","time","00:12:24.860","fde32c9","00:12:21.020","0dcc61a","00:12:15.900","bd30f38","00:12:11.840",{"_203":252,"_26":234,"_235":253,"_237":254},{"_203":250,"_26":234,"_235":251,"_237":241},{"_203":247,"_26":234,"_235":248,"_237":249},{"_203":245,"_26":234,"_235":246,"_237":241},{"_203":242,"_26":234,"_235":243,"_237":244},{"_203":239,"_26":234,"_235":240,"_237":241},{"_203":233,"_26":234,"_235":236,"_237":238},"ef8786a0d33f","span","marks",[216],"text","Thanks for having me.","1f9a6c48aa61",[]," ","c3c3b2fea857",[220],"And then my email is sarah@letta.com if you have any questions.","a5ca8649cc99",[],"49f1da7e096d",[222],"In the meantime, if you also want to try out Letta, you can just go to letta.com/agent or chat.letta.com.","12e1ba80bc7c",[],"11cd7099ab85",[224],"So this is a benchmark that we'll have out probably, hopefully in a few weeks.","f3ac177d612f",[270,271,272,273,274,275,276],[258,259,260,261],{"_203":268,"_26":217,"_218":269},{"_203":266,"_26":217,"_218":267},{"_203":264,"_26":217,"_218":265},{"_203":262,"_26":217,"_218":263},"cff28d6","00:12:04.640","f8b97f0","00:11:58.320","86db953","00:11:52.320","2dfee7e","00:11:47.380",{"_203":292,"_26":234,"_235":293,"_237":294},{"_203":290,"_26":234,"_235":291,"_237":241},{"_203":287,"_26":234,"_235":288,"_237":289},{"_203":285,"_26":234,"_235":286,"_237":241},{"_203":282,"_26":234,"_235":283,"_237":284},{"_203":280,"_26":234,"_235":281,"_237":241},{"_203":277,"_26":234,"_235":278,"_237":279},"5c19e3d7ec8a",[262],"That goes beyond just can the agent recall something, which I think is very limited.","625b48f76d83",[],"05a2e12c9864",[264],"So this actually gives us an objective way to measure how good is the learned context or memory of an agent.","24438c84b990",[],"b81b3219c991",[266],"Red is maybe more efficient, but then has a bit less reward.","4f0ab09fedd4",[],"c1db0a15a564",[268],"So blue will improve it even more, but maybe has some higher token usage.","02bc485953ad",[307,308,309,310,311],[298,299,300],{"_203":305,"_26":217,"_218":306},{"_203":303,"_26":217,"_218":304},{"_203":301,"_26":217,"_218":302},"c6e6680","00:11:41.540","9c4c2b8","00:11:37.700","91096b4","00:11:31.700",{"_203":322,"_26":234,"_235":323,"_237":324},{"_203":320,"_26":234,"_235":321,"_237":241},{"_203":317,"_26":234,"_235":318,"_237":319},{"_203":315,"_26":234,"_235":316,"_237":241},{"_203":312,"_26":234,"_235":313,"_237":314},"4031dce1b4e3",[301],"And both of these kind of to varying degrees will actually improve the agent's performance.","877945ea8443",[],"70385d5a6c57",[303],"So we have two different approaches, algorithm one, algorithm two.","84a67122c0f2",[],"be2b0a8f41fd",[305],"So we have this curve on the bottom, and then we can now measure basically different approaches to forming memory.","c9b381fe2b1d",[337,338,339,340,341],[328,329,330],{"_203":335,"_26":217,"_218":336},{"_203":333,"_26":217,"_218":334},{"_203":331,"_26":217,"_218":332},"bb98283","00:11:28.240","c423eb4","00:11:25.280","fed5d9f","00:11:20.400",{"_203":352,"_26":234,"_235":353,"_237":354},{"_203":350,"_26":234,"_235":351,"_237":241},{"_203":347,"_26":234,"_235":348,"_237":349},{"_203":345,"_26":234,"_235":346,"_237":241},{"_203":342,"_26":234,"_235":343,"_237":344},"bc6b4af7c658",[331],"So higher reasoning does better, but uses more tokens.","4be662bd1ed8",[],"0d843ac24655",[333],"So low, medium, high is just the reasoning levels.","5802354d508a",[],"f24e58eb9bea",[335],"And so you can see here, basically the gray is having no memory.","7e845990e91f",[364,365,366],[358,359],{"_203":362,"_26":217,"_218":363},{"_203":360,"_26":217,"_218":361},"8569b38","00:11:14.100","78f679b","00:11:09.779",{"_203":372,"_26":234,"_235":373,"_237":374},{"_203":370,"_26":234,"_235":371,"_237":241},{"_203":367,"_26":234,"_235":368,"_237":369},"56330b733308",[360],"The benchmark itself is actually an adaptation of existing benchmarks to require memory to solve tasks.","f79dd37a2810",[],"5f740ff8f7d6",[362],"And so these are some early results that we have from this benchmark that we're working on.","2cda21cdb371",[385,386,387],[379,380],"blockquote",{"_203":383,"_26":217,"_218":384},{"_203":381,"_26":217,"_218":382},"f1ce017","00:11:06.399","57e0bbd","00:11:03.040",{"_203":393,"_26":234,"_235":394,"_237":395},{"_203":391,"_26":234,"_235":392,"_237":241},{"_203":388,"_26":234,"_235":389,"_237":390},"f3b651c85519",[381],"So I think these are actually very, very different things.","6a8c65cc28b3",[],"bdff83670903",[383],"Did it actually learn? Not just, did it remember the task?","5d3b426778c0",[411,412,413,414,415,416,417],[399,400,401,402],{"_203":409,"_26":217,"_218":410},{"_203":407,"_26":217,"_218":408},{"_203":405,"_26":217,"_218":406},{"_203":403,"_26":217,"_218":404},"2dae3b7","00:10:54.060","8f0ff67","00:10:50.620","085e03d","00:10:45.880","4c3223f","00:10:36.040",{"_203":433,"_26":234,"_235":434,"_237":435},{"_203":431,"_26":234,"_235":432,"_237":241},{"_203":428,"_26":234,"_235":429,"_237":430},{"_203":426,"_26":234,"_235":427,"_237":241},{"_203":423,"_26":234,"_235":424,"_237":425},{"_203":421,"_26":234,"_235":422,"_237":241},{"_203":418,"_26":234,"_235":419,"_237":420},"a8f8186c5ad4",[403],"works or not. So this benchmark that we're currently working on is actually much more so about, is the agent actually better at the task?","73e8644342f8",[],"b451c7b39d53",[405],"if the recall even","ae9181d842e7",[],"b0d3bf3f392b",[407],"I think at this point recall is just how fast can you recall something, not necessarily","f7ea3156be18",[],"b60dec4bf4b0",[409],"So this is very, very different from other memory benchmarks, if you might have seen them, which I think primarily tend to focus on recall, which in my view is mostly a solved problem.","94aec7e25e12",[445,446,447],[439,440],{"_203":443,"_26":217,"_218":444},{"_203":441,"_26":217,"_218":442},"112bbb9","00:10:30.620","4125bb7","00:10:25.260",{"_203":453,"_26":234,"_235":454,"_237":455},{"_203":451,"_26":234,"_235":452,"_237":241},{"_203":448,"_26":234,"_235":449,"_237":450},"77ba8f318eb7",[441],"So take some end task and measure the efficiency and capability on that task.","237969edd6d1",[],"97670f22b6b5",[443],"And so I think the ideal way to measure this is to actually measure this in a verifiable way.","457f5244c83b",[465,466,467],[459,460],{"_203":463,"_26":217,"_218":464},{"_203":461,"_26":217,"_218":462},"d59a0d5","00:10:19.100","5458f39","00:10:07.180",{"_203":473,"_26":234,"_235":474,"_237":475},{"_203":471,"_26":234,"_235":472,"_237":241},{"_203":468,"_26":234,"_235":469,"_237":470},"aa2d7ab4c9cf",[461],"So something that's kind of context meant for the agent to adapt itself.","3611ade98329",[],"cde81e445f69",[463],"So I think to be a bit more precise, I think the way I would define the agent learning or formation of memory is basically translating experience or potentially other context into learned context.","69cb0a78673f",[488,489,490,491,492],[479,480,481],{"_203":486,"_26":217,"_218":487},{"_203":484,"_26":217,"_218":485},{"_203":482,"_26":217,"_218":483},"4a57daa","00:10:00.060","9564207","00:09:54.660","87fce54","00:09:47.460",{"_203":503,"_26":234,"_235":504,"_237":505},{"_203":501,"_26":234,"_235":502,"_237":241},{"_203":498,"_26":234,"_235":499,"_237":500},{"_203":496,"_26":234,"_235":497,"_237":241},{"_203":493,"_26":234,"_235":494,"_237":495},"ae89fc9767a3",[482],"So how good is the memory or the learned context based off of something that's more objective?","f92c1cf15ee3",[],"3b8155c710b9",[484],"But ideally instead of having these opinions, we can just evaluate things objectively.","fd8eca3e1c21",[],"209add9e461d",[486],"So I think some of the examples I gave here, I am kind of imposing my opinion about what should context for an agent look like.","5e12c01ef615",[510],[],"h2",{"_203":511,"_26":234,"_235":512,"_237":513},"5e8c6ae3beeb",[],"Measuring What Agents Learn","a0be9a9ef86e",[523,524,525],[517,518],{"_203":521,"_26":217,"_218":522},{"_203":519,"_26":217,"_218":520},"5f7d04e","00:09:35.740","b051442","00:09:28.540",{"_203":531,"_26":234,"_235":532,"_237":533},{"_203":529,"_26":234,"_235":530,"_237":241},{"_203":526,"_26":234,"_235":527,"_237":528},"ce633d529eab",[519],"This is the type of prompt you're more likely to write and it's probably gonna work a lot better, not the thing on the left, which is a series of extremely specific corrections.","828a2686c875",[],"9fa6f90de13b",[521],"So this is much more like, actually, if you were prompting an agent as a human to kind of write emails in the way that you expect them to.","07d23529bc3b",[549,550,551,552,553,554,555],[537,538,539,540],{"_203":547,"_26":217,"_218":548},{"_203":545,"_26":217,"_218":546},{"_203":543,"_26":217,"_218":544},{"_203":541,"_26":217,"_218":542},"c2db951","00:09:24.459","03cac22","00:09:19.080","450c002","00:09:11.959","6668e50","00:09:04.800",{"_203":571,"_26":234,"_235":572,"_237":573},{"_203":569,"_26":234,"_235":570,"_237":241},{"_203":566,"_26":234,"_235":567,"_237":568},{"_203":564,"_26":234,"_235":565,"_237":241},{"_203":561,"_26":234,"_235":562,"_237":563},{"_203":559,"_26":234,"_235":560,"_237":241},{"_203":556,"_26":234,"_235":557,"_237":558},"b73065773772",[541],"When multiple things are owned to the same person, combine them into a single message.","d0d6f820f3b0",[],"10872a6046fc",[543],"One email per person, not one email per topic, which you'd think is obvious.","01fb691e0338",[],"04a42c6d6014",[545],"So for example, rewriting: prose only, no bullet points, numbered lists, bold labels, etc.","887509483161",[],"1583efb613cc",[547],"And an example of a better memory is actually rewriting the guidance in the prompting to adapt to the feedback that's been given.","32b5c0f643da",[589,590,591,592,593,594,595],[577,578,579,580],{"_203":587,"_26":217,"_218":588},{"_203":585,"_26":217,"_218":586},{"_203":583,"_26":217,"_218":584},{"_203":581,"_26":217,"_218":582},"1a87705","00:09:01.120","34fe21d","00:08:53.880","3d4fc05","00:08:50.920","235b9a2","00:08:45.080",{"_203":611,"_26":234,"_235":612,"_237":613},{"_203":609,"_26":234,"_235":610,"_237":241},{"_203":606,"_26":234,"_235":607,"_237":608},{"_203":604,"_26":234,"_235":605,"_237":241},{"_203":601,"_26":234,"_235":602,"_237":603},{"_203":599,"_26":234,"_235":600,"_237":241},{"_203":596,"_26":234,"_235":597,"_237":598},"839f1da0ea49",[581],"Just all these random rules and event logs.","ba24dfb3bc56",[],"630c8c7b6f69",[583],"And obviously if you have many, many of these types of edits being made over time, the context of the agent is gonna look chaotic, right?","5fc7a6062ef2",[],"8caa8b7e5c42",[585],"So that's not gonna be that helpful for the future.","c242a26e9533",[],"0db96fd18182",[587],"It's this extremely precise, \"I made this exact mistake on this date.\"","8a24ffba06f7",[626,627,628,629,630],[617,618,619],{"_203":624,"_26":217,"_218":625},{"_203":622,"_26":217,"_218":623},{"_203":620,"_26":217,"_218":621},"765f764","00:08:37.120","7140773","00:08:33.440","3854259","00:08:29.759",{"_203":641,"_26":234,"_235":642,"_237":643},{"_203":639,"_26":234,"_235":640,"_237":241},{"_203":636,"_26":234,"_235":637,"_237":638},{"_203":634,"_26":234,"_235":635,"_237":241},{"_203":631,"_26":234,"_235":632,"_237":633},"239da7944c01",[620],"So the agent just literally writes, \"Correction, 2026-7-24, Marisa asks,\" blah, blah, blah.","4ef5221b78d8",[],"9c592e1131fd",[622],"So on the left we see an example of a very bad memory edit.","57ee73eeb20f",[],"a10c85fb77fa",[624],"So this is some user feedback that we expect the agent to adapt to.","e34118a5beee",[659,660,661,662,663,664,665],[647,648,649,650],{"_203":657,"_26":217,"_218":658},{"_203":655,"_26":217,"_218":656},{"_203":653,"_26":217,"_218":654},{"_203":651,"_26":217,"_218":652},"df44ed6","00:08:25.500","65a2edb","00:08:22.300","d7b10e9","00:08:13.500","8a40477","00:08:06.639",{"_203":681,"_26":234,"_235":682,"_237":683},{"_203":679,"_26":234,"_235":680,"_237":241},{"_203":676,"_26":234,"_235":677,"_237":678},{"_203":674,"_26":234,"_235":675,"_237":241},{"_203":671,"_26":234,"_235":672,"_237":673},{"_203":669,"_26":234,"_235":670,"_237":241},{"_203":666,"_26":234,"_235":667,"_237":668},"1572dc87704d",[651],"Just send the invites separately. Also stop sending so many individual emails, just combine them.\"","7d8253516de1",[],"d17367a00e53",[653],"Do not write a bunch of AI-like bullet points or attach the invites.","ba66956a95d6",[],"14670faa4c50",[655],"So this is an example from an experiment we did where there's kind of the simulated user, where the user says, \"Everyone thinks my email is AI now.","45d07f0ca4f7",[],"c0687faf1a2d",[657],"Another very specific example of this is kind of having good edits being made to your memory.","2aaf87ea47cb",[693,694,695],[687,688],{"_203":691,"_26":217,"_218":692},{"_203":689,"_26":217,"_218":690},"76f9974","00:07:59.560","f7c044b","00:07:52.840",{"_203":701,"_26":234,"_235":702,"_237":703},{"_203":699,"_26":234,"_235":700,"_237":241},{"_203":696,"_26":234,"_235":697,"_237":698},"8ab5d430155c",[689],"Definitely not something that you would generally want to see in your agent.","42311b7bc2de",[],"f93962e08298",[691],"And I don't have an example here, but if you open up some of the files themselves, they're very messy, very chaotic.","d04d99bb5cec",[722,723,724,725,726,727,728,729,730],[707,708,709,710,711],{"_203":720,"_26":217,"_218":721},{"_203":718,"_26":217,"_218":719},{"_203":716,"_26":217,"_218":717},{"_203":714,"_26":217,"_218":715},{"_203":712,"_26":217,"_218":713},"b13151b","00:07:50.000","600b7e4","00:07:46.800","5a7dd1b","00:07:39.360","a132923","00:07:34.480","c5b77a9","00:07:31.360",{"_203":751,"_26":234,"_235":752,"_237":753},{"_203":749,"_26":234,"_235":750,"_237":241},{"_203":746,"_26":234,"_235":747,"_237":748},{"_203":744,"_26":234,"_235":745,"_237":241},{"_203":741,"_26":234,"_235":742,"_237":743},{"_203":739,"_26":234,"_235":740,"_237":241},{"_203":736,"_26":234,"_235":737,"_237":738},{"_203":734,"_26":234,"_235":735,"_237":241},{"_203":731,"_26":234,"_235":732,"_237":733},"e279a22e551d",[712],"It doesn't have the right tools for progressive disclosure.","178c931e50fd",[],"902b8c9c5252",[714],"since it just is basically overloaded with context.","4830d1673de9",[],"a4b8c886f87c",[716],"And so at this point, the likelihood that the agent is actually using any of these files in gray is extremely low,","f03eed5d8ff3",[],"bf83e5885fbe",[718],"And it just kept on creating more and more files, creating a lot of sprawl.","209a8355cafd",[],"fd33a9dd4665",[720],"And this agent was running for, I think, a few months.","5515c98a4a1e",[766,767,768,769,770],[757,758,759],{"_203":764,"_26":217,"_218":765},{"_203":762,"_26":217,"_218":763},{"_203":760,"_26":217,"_218":761},"8c12643","00:07:27.960","b350ee5","00:07:21.880","72565f4","00:07:18.920",{"_203":781,"_26":234,"_235":782,"_237":783},{"_203":779,"_26":234,"_235":780,"_237":241},{"_203":776,"_26":234,"_235":777,"_237":778},{"_203":774,"_26":234,"_235":775,"_237":241},{"_203":771,"_26":234,"_235":772,"_237":773},"27b0f0fc27ba",[760],"The yellow is skills, and the gray is external context.","856e2015cf40",[],"5a9311b41f7c",[762],"This is not the best visualization, but the blue is basically representing files that are pinned to the system prompt.","b477aa0992aa",[],"5018caf24dc1",[764],"So, I mean, this is one example of an agent we had.","d9b9d6ef335c",[796,797,798,799,800],[787,788,789],{"_203":794,"_26":217,"_218":795},{"_203":792,"_26":217,"_218":793},{"_203":790,"_26":217,"_218":791},"05c6ebe","00:07:09.560","0e51993","00:07:05.400","5d904e4","00:06:54.300",{"_203":811,"_26":234,"_235":812,"_237":813},{"_203":809,"_26":234,"_235":810,"_237":241},{"_203":806,"_26":234,"_235":807,"_237":808},{"_203":804,"_26":234,"_235":805,"_237":241},{"_203":801,"_26":234,"_235":802,"_237":803},"d25212607bfe",[790],"So it's very important to keep context windows clean and kind of structure them well and avoid this rot and bloat problem.","7590a80f376a",[],"72e889c14f97",[792],"the less the agent is sort of able to effectively use everything in context.","67aab80575e1",[],"40afc01d8f50",[794],"And I think there's been some papers that show that the more stuff you put into context, for example, Lost in the Middle, or I think Chroma had this blog post on context rot,","8bdaf971965e",[823,824,825],[817,818],{"_203":821,"_26":217,"_218":822},{"_203":819,"_26":217,"_218":820},"e0882cd","00:06:51.180","1c8a0a6","00:06:47.740",{"_203":831,"_26":234,"_235":832,"_237":833},{"_203":829,"_26":234,"_235":830,"_237":241},{"_203":826,"_26":234,"_235":827,"_237":828},"01d1571de70d",[819],"I think the demands of context windows are only increasing.","dc9cf4a1b1a7",[],"397b38b3a166",[821],"So context windows are very, very precious, even though they're getting larger.","adf564b4e668",[846,847,848,849,850],[837,838,839],{"_203":844,"_26":217,"_218":845},{"_203":842,"_26":217,"_218":843},{"_203":840,"_26":217,"_218":841},"7d9b9f3","00:06:42.919","023dadb","00:06:33.840","ed3c864","00:06:30.080",{"_203":861,"_26":234,"_235":862,"_237":863},{"_203":859,"_26":234,"_235":860,"_237":241},{"_203":856,"_26":234,"_235":857,"_237":858},{"_203":854,"_26":234,"_235":855,"_237":241},{"_203":851,"_26":234,"_235":852,"_237":853},"e11bb5a297cc",[840],"And then I think probably the biggest issue of all these is context rot and bloat.","97cbc713e5e6",[],"c5c798cfea97",[842],"So agents will kind of take away very general, not very informative lessons from experience that doesn't actually help it in the future.","294c7a39ef1d",[],"18a3e5b529f1",[844],"The other side of that same coin is basically oversummarization.","c64bda362ac2",[873,874,875],[867,868],{"_203":871,"_26":217,"_218":872},{"_203":869,"_26":217,"_218":870},"aa0782e","00:06:21.140","8ffc2d7","00:06:17.860",{"_203":881,"_26":234,"_235":882,"_237":883},{"_203":879,"_26":234,"_235":880,"_237":241},{"_203":876,"_26":234,"_235":877,"_237":878},"dce1de25e39b",[869],"This is a problem with both token space learning and weight space learning. Basically, agents will tend to overfit their context to specific mistakes.","d1b9bd772536",[],"1aa834c80825",[871],"So one is obviously a failure to generalize.","c289387194db",[893,894,895],[887,888],{"_203":891,"_26":217,"_218":892},{"_203":889,"_26":217,"_218":890},"b9343a0","00:06:08.740","b6384a4","00:06:03.200",{"_203":901,"_26":234,"_235":902,"_237":903},{"_203":899,"_26":234,"_235":900,"_237":241},{"_203":896,"_26":234,"_235":897,"_237":898},"a4a707a30084",[889],"kind of learned context from experience, but there's actually a lot of challenges with making this work really well, which I think is a big part of why it's not super commonly done right now.","92d47de18c45",[],"d8b1e1ee6b18",[891],"And so I think it would be great if agents could actually generate all this","99b0b6c5a096",[907],[],{"_203":908,"_26":234,"_235":909,"_237":910},"37ef1b007957",[],"What Breaks in Practice","85de981eba88",[923,924,925,926,927],[914,915,916],{"_203":921,"_26":217,"_218":922},{"_203":919,"_26":217,"_218":920},{"_203":917,"_26":217,"_218":918},"06a7639","00:05:59.440","1a84d0a","00:05:52.600","25f1478","00:05:46.680",{"_203":938,"_26":234,"_235":939,"_237":940},{"_203":936,"_26":234,"_235":937,"_237":241},{"_203":933,"_26":234,"_235":934,"_237":935},{"_203":931,"_26":234,"_235":932,"_237":241},{"_203":928,"_26":234,"_235":929,"_237":930},"6dbf06206768",[917],"new adaptations.","9de474e6d791",[],"a252b57370af",[919],"But these are basically programmatic modifications to the harness so that you're actually changing the harness behavior to encode","8537ab39fa92",[],"13f1b05a6c0f",[921],"So they're also known as extensions in Pi. I think Claude Code recently also came out with mods.","88f9ef03c33a",[953,954,955,956,957],[944,945,946],{"_203":951,"_26":217,"_218":952},{"_203":949,"_26":217,"_218":950},{"_203":947,"_26":217,"_218":948},"cf63754","00:05:45.160","9368dea","00:05:40.120","3ad4214","00:05:34.600",{"_203":968,"_26":234,"_235":969,"_237":970},{"_203":966,"_26":234,"_235":967,"_237":241},{"_203":963,"_26":234,"_235":964,"_237":965},{"_203":961,"_26":234,"_235":962,"_237":241},{"_203":958,"_26":234,"_235":959,"_237":960},"cc5b9549fafa",[947],"And then you can also have mods.","1d6cb3fc0f37",[],"be6cc98f93bf",[949],"You can obviously also have them rewrite their own memory or their own external context.","762e172c7553",[],"1b3a9ac0d47d",[951],"But then there's also prompts. So you can actually have agents rewrite their own prompting over time.","362010893905",[986,987,988,989,990,991,992],[974,975,976,977],{"_203":984,"_26":217,"_218":985},{"_203":982,"_26":217,"_218":983},{"_203":980,"_26":217,"_218":981},{"_203":978,"_26":217,"_218":979},"c4f13be","00:05:30.680","f63d141","00:05:27.320","c579d15","00:05:24.040","317ae42","00:05:21.400",{"_203":1008,"_26":234,"_235":1009,"_237":1010},{"_203":1006,"_26":234,"_235":1007,"_237":241},{"_203":1003,"_26":234,"_235":1004,"_237":1005},{"_203":1001,"_26":234,"_235":1002,"_237":241},{"_203":998,"_26":234,"_235":999,"_237":1000},{"_203":996,"_26":234,"_235":997,"_237":241},{"_203":993,"_26":234,"_235":994,"_237":995},"e4d793ad6544",[978],"That's probably the most commonly rewritten form of memory today for agents.","e9ccc9e03c87",[],"a7dbae654d1b",[980],"So, again, probably the best-known one is skills.","5b13b480774e",[],"b2493f055697",[982],"I think there's a couple of different types of context that agents have.","49980b171995",[],"106b81a51d04",[984],"So what does continual learning in token space mean?","1b2344112424",[1023,1024,1025,1026,1027],[1014,1015,1016],{"_203":1021,"_26":217,"_218":1022},{"_203":1019,"_26":217,"_218":1020},{"_203":1017,"_26":217,"_218":1018},"baa819b","00:05:12.020","344c46c","00:05:06.580","a47d5b5","00:05:03.380",{"_203":1038,"_26":234,"_235":1039,"_237":1040},{"_203":1036,"_26":234,"_235":1037,"_237":241},{"_203":1033,"_26":234,"_235":1034,"_237":1035},{"_203":1031,"_26":234,"_235":1032,"_237":241},{"_203":1028,"_26":234,"_235":1029,"_237":1030},"cf918d02b07a",[1017],"And so we kind of at a high level talk about this in terms of generating learned context, or continual learning in token space.","a76423a13717",[],"5afc617bddc8",[1019],"and also rewriting components of the harness, assuming that the harness exposes interfaces for that.","a698b1f30f3e",[],"5124639e837d",[1021],"So this can both be through rewriting its own prompting","2c249bdda2a6",[1050,1051,1052],[1044,1045],{"_203":1048,"_26":217,"_218":1049},{"_203":1046,"_26":217,"_218":1047},"0386089","00:04:54.900","011c1b9","00:04:44.660",{"_203":1058,"_26":234,"_235":1059,"_237":1060},{"_203":1056,"_26":234,"_235":1057,"_237":241},{"_203":1053,"_26":234,"_235":1054,"_237":1055},"19fc515f80f8",[1046],"So avoiding redundant work, knowing what to do next time to be faster and better, not make the same mistakes, and then also adapting their own behavior.","5e9ce1ef6b9b",[],"ecf175423ef3",[1048],"But, I mean, beyond just remembering, I think what you can actually do with context if you use it effectively is you can allow agents to actually improve their efficiency over time.","4c19d1d40461",[1070,1071,1072],[1064,1065],{"_203":1068,"_26":217,"_218":1069},{"_203":1066,"_26":217,"_218":1067},"382e2b6","00:04:37.760","b98aac6","00:04:30.240",{"_203":1078,"_26":234,"_235":1079,"_237":1080},{"_203":1076,"_26":234,"_235":1077,"_237":241},{"_203":1073,"_26":234,"_235":1074,"_237":1075},"12a8e66cbd26",[1066],"Or I think memory, usually people are kind of talking about it in the sense of just recalling the past, which I think is a very small portion of it.","97d18fc0ba0a",[],"9b43bc0dd4f5",[1068],"And so coming back to this idea of context and memory, I think right now memory is not really leveraged very well.","eed8b20d411d",[1084],[],{"_203":1085,"_26":234,"_235":1086,"_237":1087},"f09f32fed089",[],"Continual Learning in Token Space","570be557a223",[1097,1098,1099],[1091,1092],{"_203":1095,"_26":217,"_218":1096},{"_203":1093,"_26":217,"_218":1094},"c681a03","00:04:20.900","63635c9","00:04:15.700",{"_203":1105,"_26":234,"_235":1106,"_237":1107},{"_203":1103,"_26":234,"_235":1104,"_237":241},{"_203":1100,"_26":234,"_235":1101,"_237":1102},"a34e26c8acf8",[1093],"So it's very useful to have an SDK like this where you can have agents themselves orchestrate other agents through writing these types of programs.","b66a7d0909de",[],"573bbfde2e5d",[1095],"And then also I think more and more we're going to see agents programmatically managing other agents.","876e5cb2bbd8",[1117,1118,1119],[1111,1112],{"_203":1115,"_26":217,"_218":1116},{"_203":1113,"_26":217,"_218":1114},"b1daf90","00:04:12.420","3c79e0b","00:04:10.500",{"_203":1125,"_26":234,"_235":1126,"_237":1127},{"_203":1123,"_26":234,"_235":1124,"_237":241},{"_203":1120,"_26":234,"_235":1121,"_237":1122},"cbdb6f0213bb",[1113],"First of all, you can just build your own custom UIs for Letta agents.","3571618369e1",[],"06f22919daf4",[1115],"I think this is useful for a number of reasons.","3369e4e812d5",[1140,1141,1142,1143,1144],[1131,1132,1133],{"_203":1138,"_26":217,"_218":1139},{"_203":1136,"_26":217,"_218":1137},{"_203":1134,"_26":217,"_218":1135},"e1c7c8f","00:04:06.660","d9221f2","00:04:02.379","b534f5a","00:03:57.980",{"_203":1155,"_26":234,"_235":1156,"_237":1157},{"_203":1153,"_26":234,"_235":1154,"_237":241},{"_203":1150,"_26":234,"_235":1151,"_237":1152},{"_203":1148,"_26":234,"_235":1149,"_237":241},{"_203":1145,"_26":234,"_235":1146,"_237":1147},"0ccb27cc7e23",[1134],"that's running on some specified location and programmatically interact with that.","2db07c1392ac",[],"2b716d9aed01",[1136],"So you can basically interact with some specific agent ID","e9583a781fbc",[],"ae9fb940a05d",[1138],"So this is a quick snippet of what the Letta agents SDK looks like.","3f87e0b8d7bd",[1167,1168,1169],[1161,1162],{"_203":1165,"_26":217,"_218":1166},{"_203":1163,"_26":217,"_218":1164},"b5d8279","00:03:51.280","625b0f3","00:03:45.280",{"_203":1175,"_26":234,"_235":1176,"_237":1177},{"_203":1173,"_26":234,"_235":1174,"_237":241},{"_203":1170,"_26":234,"_235":1171,"_237":1172},"e938c7d86a1e",[1163],"stateful agents is that they exist on a server, and so you can actually interact with them programmatically through an API.","f486d13f6faa",[],"d45b0e5216c9",[1165],"But usually they're more professional. And then also on top of that, another thing about","ba1c00593a4c",[1190,1191,1192,1193,1194],[1181,1182,1183],{"_203":1188,"_26":217,"_218":1189},{"_203":1186,"_26":217,"_218":1187},{"_203":1184,"_26":217,"_218":1185},"539faf7","00:03:40.480","0aa6c60","00:03:32.920","77e9edd","00:03:28.520",{"_203":1205,"_26":234,"_235":1206,"_237":1207},{"_203":1203,"_26":234,"_235":1204,"_237":241},{"_203":1200,"_26":234,"_235":1201,"_237":1202},{"_203":1198,"_26":234,"_235":1199,"_237":241},{"_203":1195,"_26":234,"_235":1196,"_237":1197},"addbbfc3e2cd",[1184],"So you'll see kind of funny interactions like that if you introduce your agents into your Slack.","ffd942d0f638",[],"46c2363b067a",[1186],"So this was an example of some agents having an interaction where one agent told another one to stand down on a PR.","fe894aa93d94",[],"55a9e4332506",[1188],"And then these agents, you can embed them into Slack or other applications.","ad26715ec621",[1220,1221,1222,1223,1224],[1211,1212,1213],{"_203":1218,"_26":217,"_218":1219},{"_203":1216,"_26":217,"_218":1217},{"_203":1214,"_26":217,"_218":1215},"934c940","00:03:19.140","fc8451c","00:03:12.180","f6000f7","00:03:08.420",{"_203":1235,"_26":234,"_235":1236,"_237":1237},{"_203":1233,"_26":234,"_235":1234,"_237":241},{"_203":1230,"_26":234,"_235":1231,"_237":1232},{"_203":1228,"_26":234,"_235":1229,"_237":241},{"_203":1225,"_26":234,"_235":1226,"_237":1227},"89f5e49c091e",[1214],"All the agents are basically custom, so you can create different custom agents that can interact across your team or work across agents as well.","abb804ed11b5",[],"79ab0b81287c",[1216],"We basically have a desktop app and web app where you can run agents in cloud sandboxes or your local machine.","ae26f114ab4e",[],"231f4395af5a",[1218],"So this is kind of a quick glimpse of our product.","876762f5420f",[1247,1248,1249],[1241,1242],{"_203":1245,"_26":217,"_218":1246},{"_203":1243,"_26":217,"_218":1244},"e84e3cb","00:03:01.680","8b814f6","00:02:59.440",{"_203":1255,"_26":234,"_235":1256,"_237":1257},{"_203":1253,"_26":234,"_235":1254,"_237":241},{"_203":1250,"_26":234,"_235":1251,"_237":1252},"81abcee84d07",[1243],"You have this concept of an agent identity, and it's always the same agent with the same knowledge wherever you go.","360a7e9e40c4",[],"5dd038b24f18",[1245],"So this is something that we support in Letta.","0c37a93f56b8",[1267,1268,1269],[1261,1262],{"_203":1265,"_26":217,"_218":1266},{"_203":1263,"_26":217,"_218":1264},"9e997b2","00:02:57.440","f0aa948","00:02:51.600",{"_203":1275,"_26":234,"_235":1276,"_237":1277},{"_203":1273,"_26":234,"_235":1274,"_237":241},{"_203":1270,"_26":234,"_235":1271,"_237":1272},"bdef0047a1e4",[1263],"But you have all these same interfaces.","acdab190c4be",[],"74d08eaf93f3",[1265],"It's gonna know about all your past interactions, it's gonna have all the adaptations it might have introduced before.","b9adf2383176",[1290,1291,1292,1293,1294],[1281,1282,1283],{"_203":1288,"_26":217,"_218":1289},{"_203":1286,"_26":217,"_218":1287},{"_203":1284,"_26":217,"_218":1285},"20d8675","00:02:44.960","f31274d","00:02:42.300","9879183","00:02:37.180",{"_203":1305,"_26":234,"_235":1306,"_237":1307},{"_203":1303,"_26":234,"_235":1304,"_237":241},{"_203":1300,"_26":234,"_235":1301,"_237":1302},{"_203":1298,"_26":234,"_235":1299,"_237":241},{"_203":1295,"_26":234,"_235":1296,"_237":1297},"74b29ee7bd20",[1284],"that you're interacting with from a CLI, from Slack, from iMessage, from some web app, and it's still the same thing.","f41f468a72ec",[],"cac5f7db28cb",[1286],"So you can have, again, the same agent identity","cbc85d9944c1",[],"59fa983f66cf",[1288],"And on top of agents being able to be multi-machine, it also means that you can have agents become multi-interface.","fdb8446da6e7",[1311],[],{"_203":1312,"_26":234,"_235":1313,"_237":1314},"cb575a8c0080",[],"One Agent, Many Interfaces","f6fcf7bb029b",[1330,1331,1332,1333,1334,1335,1336],[1318,1319,1320,1321],{"_203":1328,"_26":217,"_218":1329},{"_203":1326,"_26":217,"_218":1327},{"_203":1324,"_26":217,"_218":1325},{"_203":1322,"_26":217,"_218":1323},"0fa08de","00:02:28.540","bddc0b4","00:02:20.920","6ecaec7","00:02:14.360","e5d1cd2","00:02:10.060",{"_203":1352,"_26":234,"_235":1353,"_237":1354},{"_203":1350,"_26":234,"_235":1351,"_237":241},{"_203":1347,"_26":234,"_235":1348,"_237":1349},{"_203":1345,"_26":234,"_235":1346,"_237":241},{"_203":1342,"_26":234,"_235":1343,"_237":1344},{"_203":1340,"_26":234,"_235":1341,"_237":241},{"_203":1337,"_26":234,"_235":1338,"_237":1339},"e19e8f50569f",[1322],"We have a blog post on context repositories that talks a little bit more about how you can have these managed context repositories to build these types of agents.","658ca44f1a79",[],"b07a66c444b7",[1324],"so that you're actually interacting with this persistent thing rather than just spawning these short-lived sessions.","09121501eb42",[],"7a43a79e2e56",[1326],"your own laptop, some Mac Mini, and it's still kind of like the same agent that's tied to the same context and identity,","c3dc541fa8e3",[],"755436bd3de2",[1328],"So you could have the same agent identity essentially working across sandboxes,","f19c38f77bcc",[1367,1368,1369,1370,1371],[1358,1359,1360],{"_203":1365,"_26":217,"_218":1366},{"_203":1363,"_26":217,"_218":1364},{"_203":1361,"_26":217,"_218":1362},"f800b92","00:02:03.740","0ef44ea","00:01:53.400","fcc92c9","00:01:42.840",{"_203":1382,"_26":234,"_235":1383,"_237":1384},{"_203":1380,"_26":234,"_235":1381,"_237":241},{"_203":1377,"_26":234,"_235":1378,"_237":1379},{"_203":1375,"_26":234,"_235":1376,"_237":241},{"_203":1372,"_26":234,"_235":1373,"_237":1374},"7b1ddd9a596b",[1361],"other external contexts, basically context that's part of the agent and is carried wherever it works.","c99b0f1f5b61",[],"e8bfcc13df42",[1363],"So this can include things like skills, which I think most people are probably familiar with, but also its own prompting, potentially programmatic modifications to the harness like mods,","c946500c7a80",[],"38e82f739a2e",[1365],"And so what this kind of looks like at a very high level is that the agent itself is tied to both all of its past conversations and then also context that it owns.","1811cc1a1fec",[1397,1398,1399,1400,1401],[1388,1389,1390],{"_203":1395,"_26":217,"_218":1396},{"_203":1393,"_26":217,"_218":1394},{"_203":1391,"_26":217,"_218":1392},"08a8e30","00:01:34.420","9b9b352","00:01:26.820","3e76843","00:01:18.119",{"_203":1412,"_26":234,"_235":1413,"_237":1414},{"_203":1410,"_26":234,"_235":1411,"_237":241},{"_203":1407,"_26":234,"_235":1408,"_237":1409},{"_203":1405,"_26":234,"_235":1406,"_237":241},{"_203":1402,"_26":234,"_235":1403,"_237":1404},"468cbe4c9cf4",[1391],"kind of like human beings are, and then also have native memory and learning from experience built into the harness.","ae955364f385",[],"9bc0aaf8164d",[1393],"So what this means is you can then have agents that are actually multi-user, multi-agent, multi-machine,","f23b4b584c6b",[],"239b69030283",[1395],"But I think what we're moving towards more and more is instead having proactive agents and agents that are actually stateful and persistent.","e6a9cee1d795",[1427,1428,1429,1430,1431],[1418,1419,1420],{"_203":1425,"_26":217,"_218":1426},{"_203":1423,"_26":217,"_218":1424},{"_203":1421,"_26":217,"_218":1422},"d28ecf7","00:01:12.119","5c67169","00:01:10.040","7df384a","00:01:05.560",{"_203":1442,"_26":234,"_235":1443,"_237":1444},{"_203":1440,"_26":234,"_235":1441,"_237":241},{"_203":1437,"_26":234,"_235":1438,"_237":1439},{"_203":1435,"_26":234,"_235":1436,"_237":241},{"_203":1432,"_26":234,"_235":1433,"_237":1434},"778ad5e4997a",[1421],"You probably don't share traces across your team or have agents use it to improve in the future.","b91dad1ec9dc",[],"3e67f6902d99",[1423],"It's never actually used for anything useful.","f2ebe8a8172c",[],"1974680f4dc1",[1425],"They might have some session state or context, but it's probably eventually deleted.","da275990a5c0",[1457,1458,1459,1460,1461],[1448,1449,1450],{"_203":1455,"_26":217,"_218":1456},{"_203":1453,"_26":217,"_218":1454},{"_203":1451,"_26":217,"_218":1452},"5b9dda3","00:00:57.860","ee4dd58","00:00:49.820","0c817a4","00:00:42.220",{"_203":1472,"_26":234,"_235":1473,"_237":1474},{"_203":1470,"_26":234,"_235":1471,"_237":241},{"_203":1467,"_26":234,"_235":1468,"_237":1469},{"_203":1465,"_26":234,"_235":1466,"_237":241},{"_203":1462,"_26":234,"_235":1463,"_237":1464},"c20053373dd1",[1451],"Because I think today most harnesses that you're probably using are just these ephemeral processes that run on your local machine.","49ae88a94588",[],"2c2cc882afaf",[1453],"They're more like a person that actually learns and has state and kind of exists outside of just ephemeral sessions.","d37c266d8fca",[],"119055934746",[1455],"And I think the core of what we are basically working on is, how do you create these persistent stateful agents that actually last forever?","1d7eb124ceff",[1478],[],{"_203":1479,"_26":234,"_235":1480,"_237":1481},"2a7bd9213a71",[],"Persistent, Stateful Agents","f37949b0b6d9",[1494,1495,1496,1497,1498],[1485,1486,1487],{"_203":1492,"_26":217,"_218":1493},{"_203":1490,"_26":217,"_218":1491},{"_203":1488,"_26":217,"_218":1489},"685950e","00:00:36.880","419886b","00:00:29.920","a8335bd","00:00:20.580",{"_203":1509,"_26":234,"_235":1510,"_237":1511},{"_203":1507,"_26":234,"_235":1508,"_237":241},{"_203":1504,"_26":234,"_235":1505,"_237":1506},{"_203":1502,"_26":234,"_235":1503,"_237":241},{"_203":1499,"_26":234,"_235":1500,"_237":1501},"3dddbc607542",[1488],"And then also things like memory as file system and also git-backed context.","be189bbc37d3",[],"301e32626a90",[1490],"So basically the idea of having agents scale compute on context offline so that they can be better for future tasks.","8c85f95ffd44",[],"2bfe2749fc29",[1492],"So some of our work has included MemGPT, which was the first example of agents that self-manage memory, also sleep-time compute, which I think is now kind of referred to as dreaming.","5af0bfcb74a8",[1521,1522,1523],[1515,1516],{"_203":1519,"_26":217,"_218":1520},{"_203":1517,"_26":217,"_218":1518},"05ae08a","00:00:14.020","afc44ae","00:00:04.320",{"_203":1529,"_26":234,"_235":1530,"_237":1531},{"_203":1527,"_26":234,"_235":1528,"_237":241},{"_203":1524,"_26":234,"_235":1525,"_237":1526},"82b5af791325",[1517],"So our company, Letta, just as a quick introduction, we are working on how to create agents that learn.","c4642da0ed87",[],"0b04b540e3d2",[1519],"Today I'm going to be talking about persistent memory systems, or how to build long-lived stateful agents that actually learn and improve over time.",{"_18":1583,"_20":1584,"_22":1585,"_26":1542,"_28":1583,"_1543":1586,"_1545":1587,"_14":1588},{"_18":1557,"_20":1578,"_22":1559,"_26":1542,"_28":1560,"_1543":1579,"_1545":1580,"_14":1581},{"_18":1570,"_20":1571,"_22":1572,"_26":1542,"_28":1573,"_1543":1574,"_1545":1575,"_14":1576},{"_18":1557,"_20":1565,"_22":1559,"_26":1542,"_28":1560,"_1543":1566,"_1545":1567,"_14":1568},{"_18":1557,"_20":1558,"_22":1559,"_26":1542,"_28":1560,"_1543":1561,"_1545":1562,"_14":1563},{"_18":1550,"_20":1551,"_22":1552,"_26":1542,"_28":1550,"_1543":1553,"_1545":1554,"_14":1555},{"_18":1539,"_20":1540,"_22":1541,"_26":1542,"_28":1539,"_1543":1544,"_1545":1546,"_14":1547},"2023-07-17T23:23:49Z","eec559a9-10e7-4405-82e4-e0736495ec36","TfMkf5gVJ7UWv1vMi5Qxor","tag","description","Artificial intelligence, including AI-powered coding assistants built on large language models.","name","AI",{"_26":14,"_1548":1549},"current","ai","2025-04-15T17:19:22Z","80867d7d-9222-41f5-a465-915aa791bc97","YlxgmKTLxstUdop23VNy4f","AI agents are autonomous systems built to independently take action on business or technical tasks.","AI Agents",{"_26":14,"_1548":1556},"ai-agents","2026-10-23T23:10:26Z","6625b640-07c5-4ac4-9324-bf13e52c5ed3","mG7vmCjaB1vEw3gqUIqrjf","2026-10-23T23:11:12Z","Agents that work toward goals over hours or days, and the infrastructure that keeps them on track.","Long-Horizon Agents",{"_26":14,"_1548":1564},"long-horizon-agents","55e6b0bf-3cf3-4fb8-b73d-a89a18180b76","Evaluating agent behavior, skills, and learning.","Agent Evals",{"_26":14,"_1548":1569},"agent-evals","2025-12-09T17:45:33Z","2a3f44e0-7a12-4730-a2e2-695a100b75be","6ACXEUwYcGP2Q5rQl5ATFI","2025-12-09T17:47:31Z","Context engineering is the process of designing the information payload entered into a prompt to improve LLMs' chances of efficiently returning an accurate response.","Context Engineering",{"_26":14,"_1548":1577},"context-engineering","1fc6fbde-4346-46ff-84e3-43a620e9a3f4","How agents store, recall, forget, and learn from past work.","Agent Memory",{"_26":14,"_1548":1582},"agent-memory","2024-09-05T17:31:35Z","017fa9be-32e2-46ee-bddf-f13275fbdf00","aeN0todGvIjkWyTywFzV6F","Pertaining to the technology underpinning AI and machine learning operations.","AI Infrastructure",{"_26":14,"_1548":1589},"ai-infrastructure",{"_203":1591,"_26":1592,"_1593":-5,"_1594":1595},"spk0sarahw","speaker","organization","person",{"_18":1596,"_20":1597,"_22":1598,"_24":1599,"_26":1594,"_28":1600,"_1601":1602,"_1603":33,"_1545":1604,"_1593":1605,"_1606":12,"_1607":1608,"_1609":1610,"_71":1611,"_14":1612,"_1613":12},"2026-09-10T17:40:49Z","c821fccf-e91d-4375-ad08-1b059abbab24","5D0tkT3XG0RBypwrpyt2Ob",{"_1645":1646},"2026-10-26T21:31:31Z","bio",[1630],"generate","Sarah Wooders",{"_18":1622,"_20":1623,"_22":1624,"_26":1593,"_28":1625,"_1603":12,"_1626":1627,"_1545":1628,"_1629":12},"partner","photo",{"_26":1617,"_77":1618},"position","Cofounder and CTO",{"_26":71,"_1615":12,"_1616":12},{"_26":14,"_1548":1614},"teamMember","sarah-wooders","nofollow","noindex","mainImage",{"_1619":1620,"_26":1621},"_ref","image-04a865f18f2fe7ebeb49fdcec095da23ce1a575b-416x416-jpg","reference","2026-09-10T17:40:01Z","7ace2f3a-9738-48b4-a9ba-a9404571ba5a","AvBrD6DdjyccEYuoF7qYYQ","2026-09-10T17:40:13Z","link","https://www.letta.com","Letta","portfolioCompany",{"_203":1631,"_26":205,"_206":1632,"_208":1633,"_210":211},"b1",[1638,1639],[1634],{"_203":1635,"_26":1626,"_1636":1637},"l1","href","https://www.linkedin.com/in/wooders/",{"_203":1643,"_26":234,"_235":1644,"_237":1604},{"_203":1640,"_26":234,"_235":1641,"_237":1642},"s2",[]," is Cofounder and CTO of Letta and co-author of MemGPT. She works on persistent memory for agents operating across code, browsers, desktops, and the cloud.","s1",[1635],"base",{"_103":1597,"_1647":1648},"rev","IBgcBIPGoUVrOhZYuqvaTY","2026-10-26T15:35:36Z","6966e1de-dcf5-4382-83d9-920290d3bd1a","C6cJy7Ex0HWbukQjnu57xX",{"_1645":1849},"2026-10-27T01:27:09Z","crucibleSkill","agent-infrastructure","The runtime substrate for AI agents: the machine an agent runs on, sandboxes and isolation, agent memory and learning, durable execution and harness design, and the performance and cost of running agents. Drawn from practitioners at exe.dev, Daytona, Letta, Glean, and others.","groupSlug","ai-and-ml","headline","The machines and memory agents run on","order",5,"skillUrl","https://github.com/heavybit/skills/releases/download/skills-latest/agent-infrastructure.zip",{"_26":14,"_1548":1655},"sources",[1811,1812,1813,1814,1815,1816,1817,1818,1819,1820,1821,1822,1823],[1670,1671,1672,1673,1674,1675,1676,1677,1678,1679,1680,1681,1682,1683,1684,1685,1686,1687],"Agent Infrastructure",{"_20":1805,"_1545":1806,"_1607":1807},{"_20":1789,"_1545":1790,"_1607":1791},{"_20":1784,"_1545":1785,"_1607":1786},{"_20":1776,"_1545":1777,"_1607":1778},{"_20":1774,"_1545":1775,"_1607":-5},{"_20":1746,"_1545":1747,"_1607":1748},{"_20":1744,"_1545":1745,"_1607":-5},{"_20":1739,"_1545":1740,"_1607":1741},{"_20":1597,"_1545":1604,"_1607":1737},{"_20":1732,"_1545":1733,"_1607":1734},{"_20":1726,"_1545":1727,"_1607":1728},{"_20":1718,"_1545":1719,"_1607":1720},{"_20":1713,"_1545":1714,"_1607":1715},{"_20":1708,"_1545":1709,"_1607":1710},{"_20":1703,"_1545":1704,"_1607":1705},{"_20":1698,"_1545":1699,"_1607":1700},{"_20":1693,"_1545":1694,"_1607":1695},{"_20":1688,"_1545":1689,"_1607":1690},"9b13877f-3c68-4b5a-bc2e-d4a9d74e6ca5","Josh Bleecher Snyder",{"_26":1617,"_77":1691},{"_1619":1692,"_26":1621},"image-665fbad69371e2c1120954174d8bffa3b0d12401-256x256-png","8c124d08-abfb-431f-93dc-95aefffe33c7","Tyler Edwards",{"_26":1617,"_77":1696},{"_1619":1697,"_26":1621},"image-063c342c38822f4ec042dd2105b20ed6bcc4f0f4-3667x3667-jpg","79e42362-591e-413b-b5ec-18945e1c6034","Pasha Sadri",{"_26":1617,"_77":1701},{"_1619":1702,"_26":1621},"image-60681c66e5f3e6d9d794aea80ee177672bd4a272-800x800-png","24f31792-7fcc-4293-9a99-103e723369fd","Caitlin Colgrove",{"_26":1617,"_77":1706},{"_1619":1707,"_26":1621},"image-9cc082fdbd8b105a8dd7a7667ae0ae9e084add82-547x365-jpg","d548e96f-977b-42d7-bfa5-86e8080d8839","Chad Metcalf",{"_26":1617,"_77":1711},{"_1619":1712,"_26":1621},"image-49a64bf4a8f917b66d31cdd25abdc7eb521f02fe-800x800-jpg","440cd747-cee9-45f5-b105-b65109870963","Tony Gentilcore",{"_26":1617,"_77":1716},{"_1619":1717,"_26":1621},"image-5682472d511a94355ba04496b4a9b8d08efcef74-200x200-jpg","person-3522","Brian Douglas",{"_26":1617,"_1721":1722,"_77":1723,"_1724":1722},"alt","Brian Douglas Headshot",{"_1619":1725,"_26":1621},"caption","image-38643ec02b6d19407e27dfb5687bb37265a6bbd5-4160x6240-jpg","155f0af2-6dee-4b64-aa91-ee84d973e629","Rob Ferguson",{"_26":1617,"_77":1729,"_1724":1730},{"_1619":1731,"_26":1621},"Rob-Ferguson-Headshot","image-9f8d86bb0bec1f8365f5f89cd83f87e90862b8e9-800x800-jpg","a655dff0-6570-4e40-bdfa-3a0fefcc4ded","Mudith Jayasekara",{"_26":1617,"_77":1735},{"_1619":1736,"_26":1621},"image-80445741bb82ca8bac8cc91d42d953917261572a-400x400-jpg",{"_26":1617,"_77":1738},{"_1619":1620,"_26":1621},"404c48b2-c631-4341-a2d3-96a8b7f955a0","Will Cory",{"_26":1617,"_77":1742},{"_1619":1743,"_26":1621},"image-d77a449a7beb6bf0593236cc43fb139aab6a2680-800x800-png","0c2cbd10-dd24-43cb-b762-a5d7e3c61ece","Mike Hostetler","7fbf3d51-584f-4e07-a1af-1b3d207c043a","Amir Zohrenejad",{"_26":1617,"_1721":1749,"_77":1750,"_1751":1752,"_1753":1754},"Amir Zohrenejad's Headshot",{"_1619":1773,"_26":1621},"crop",{"_26":1764,"_1765":1766,"_1767":1768,"_1769":1770,"_1771":1772},"hotspot",{"_26":1755,"_1756":1757,"_1758":1759,"_1760":1761,"_1762":1763},"sanity.imageHotspot","height",0.5134356192252267,"width",0.5074308563412813,"x",0.464549697141559,"y",0.30786428753072814,"sanity.imageCrop","bottom",0.4354179028566585,"left",0.21083426897091836,"right",0.28173487468780034,"top",0.051146477918114774,"image-a516037bcbe74034f1006d758dac5b252ea2d957-4800x6000-jpg","b3d34155-356e-41f9-9ea6-b18675e2d26f","Heinrich Krupp","recWJhIxYZdl8jrez","Marc Campbell",{"_26":1617,"_1721":1779,"_77":1780,"_1781":1782,"_1724":1779},"Photo of Marc Campbell",{"_1619":1783,"_26":1621},"attribution","","image-6b8bd258cf660df1894613f920bdf3aaef512e5b-306x306-jpg","294f6a90-343d-4a45-98c5-78d67d108abe","Ivan Burazin",{"_26":1617,"_77":1787},{"_1619":1788,"_26":1621},"image-39840bd1899e8995ef939c611fa59ca96a74d71b-800x800-jpg","person-3823","Joseph Ruscio",{"_26":1617,"_1721":1792,"_77":1793,"_1724":1790,"_1751":1794,"_1753":1795},"Joseph Ruscio Headshot",{"_1619":1804,"_26":1621},{"_26":1764,"_1765":1800,"_1767":1801,"_1769":1802,"_1771":1803},{"_26":1755,"_1756":1796,"_1758":1797,"_1760":1798,"_1762":1799},0.589705009748241,0.641046643838368,0.5251721653586651,0.2948525048741205,0.410294990251759,0.2046488434394811,0.1543045127221509,0,"image-6ef9c9d958990f65424c0cf26ae57978c37993a5-4800x6000-jpg","recycRa6eMq6dxTSk","David Crawshaw",{"_26":1617,"_1721":1808,"_77":1809,"_1781":1782,"_1724":1808},"Photo of David Crawshaw",{"_1619":1810,"_26":1621},"image-1c07ed5357312ea790a0cbee343450ea2e9ef74a-500x335-jpg",{"_203":1847,"_1619":1848,"_26":1621},{"_203":1845,"_1619":1846,"_26":1621},{"_203":1843,"_1619":1844,"_26":1621},{"_203":1841,"_1619":1842,"_26":1621},{"_203":1839,"_1619":1840,"_26":1621},{"_203":1837,"_1619":1838,"_26":1621},{"_203":1835,"_1619":1836,"_26":1621},{"_203":1833,"_1619":1834,"_26":1621},{"_203":1831,"_1619":1832,"_26":1621},{"_203":1830,"_1619":21,"_26":1621},{"_203":1828,"_1619":1829,"_26":1621},{"_203":1826,"_1619":1827,"_26":1621},{"_203":1824,"_1619":1825,"_26":1621},"daae494039ce4816b85c3bcb","daae4940-39ce-4816-b85c-3bcb0d1b354d","4cf0dd206e504e5491f218e6","4cf0dd20-6e50-4e54-91f2-18e60d34af37","4aee5308fa0d454c9022c1ed","4aee5308-fa0d-454c-9022-c1ed53541d2d","39c370db08bf4de5b2c27b57","18b722517b1445b2ab6cca37","18b72251-7b14-45b2-ab6c-ca370ea48cd5","7f2db4aa8e9345dba467a326","7f2db4aa-8e93-45db-a467-a326acd92cfb","f7c0c2047a254ff7968e42da","f7c0c204-7a25-4ff7-968e-42da9b0a3506","39be2ceb460040c480a9dc0c","39be2ceb-4600-40c4-80a9-dc0c7b9504dd","87ce419b74da43849e372223","87ce419b-74da-4384-9e37-22234194cec6","507fa92fa83b47998ff87523","507fa92f-a83b-4799-8ff8-7523ebaa5df4","fb22b02e08f04aa999464208","fb22b02e-08f0-4aa9-9946-4208e10a8bdf","5e7c0decaea64ee7bc087932","5e7c0dec-aea6-4ee7-bc08-793205f187b1","f5d4fea2adcc444ca4c829c6","f5d4fea2-adcc-444c-a4c8-29c60d2d838d",{"_103":1650,"_1647":1850},"VHTTFyv5E2QrBnUR21VSsp","image",{"_1619":1853,"_26":1621},"image-1c521bac60f5045ce359819db293c98ca8dc022c-1920x1080-jpg",{"_203":1855,"_26":205,"_206":1856,"_208":1857,"_210":211},"90483699479e",[1858],[],{"_203":1859,"_26":234,"_235":1860,"_237":1861},"4a7196ce18fc",[],"

Persistent Memory + What Breaks in Production\n
- Persistent, Stateful Agents
- One Agent, Many Interfaces
- Continual Learning in Token Space
- What Breaks in Practice
- Measuring What Agents Learn
\n

",{"_203":1868,"_1619":1869,"_26":1621},{"_203":1867,"_1619":1827,"_26":1621},{"_203":1865,"_1619":1866,"_26":1621},"4TabjKKXQLFu","67355535-cdb5-472f-b416-5b48b1294f09","iiAfMS8J5OLd","DAzPFc9UgeXL","27ad23dc-58b9-4366-a3b9-c73d40aedc73","items",[1925,1926,1927],"recent",[1874,1875,1876],{"_1877":1911,"_99":1912,"_1851":-5,"_1881":1913,"_53":1914,"_14":1915,"_1885":1886,"_66":1916,"_1888":1889},{"_1877":1898,"_99":1899,"_1851":-5,"_1881":1834,"_53":1900,"_14":1901,"_1885":1886,"_66":1902,"_1888":1889},{"_1877":1878,"_99":1879,"_1851":1880,"_1881":1882,"_53":1883,"_14":1884,"_1885":1886,"_66":1887,"_1888":1889},"content",[1890],17,"https://cdn.sanity.io/images/50q6fr1p/production/28801f05c737cc57c3c04d467a0a54838fc70ad0-1200x630.png","objectID","9d8377ac-25e6-47b7-8e40-22ee9a7cb43e","2026-08-27T16:38:00.000Z","understanding-code-by-treating-it-as-a-graph","templateType","default","Is the Key to Understanding Code Treating it as a Graph?","typeName","blogArticle",{"_203":1891,"_26":205,"_206":1892,"_208":1893,"_210":211},"d8cde6678d08",[1894],[],{"_203":1895,"_26":234,"_235":1896,"_237":1897},"27e674f3e7de",[],"Treating software as a graph may be the key to quickly understanding and maintaining software in the age of AI. Open-source creator Vitali Avagyan explains.",[1903],20,"2026-10-02T19:09:00.000Z","long-horizon-agents-part-2-memory-is-more-than-recall","Long-Horizon Agents, Part 2: Memory Is More than Recall",{"_203":1904,"_26":205,"_206":1905,"_208":1906,"_210":211},"82f5581705a2",[1907],[],{"_203":1908,"_26":234,"_235":1909,"_237":1910},"4283a07b60d3",[],"Agents need persistent memory to become performant for long-horizon jobs. Amir Zohrenejad surveys the state of agentic memory and covers what’s missing.",[1917],16,"616147c4-aa4e-4ab9-83ae-f20398e6871d","2026-10-20T16:48:00.000Z","how-to-orchestrate-multiple-agents-from-the-terminal","How to Orchestrate Multiple Agents from the Terminal",{"_203":1918,"_26":205,"_206":1919,"_208":1920,"_210":211},"19b8b2c958d1",[1921],[],{"_203":1922,"_26":234,"_235":1923,"_237":1924},"0cbb628f4fd2",[],"Open-source creator Gaurav Gosain explains how his work with terminal multiplexers led him to build multi-agent orchestration.",{"_1877":1958,"_99":1959,"_1851":1960,"_1881":1869,"_1931":1932,"_53":1961,"_14":1962,"_66":1963,"_1888":27},{"_1877":1944,"_99":1945,"_1851":1946,"_1881":1827,"_1931":1932,"_53":1947,"_14":1948,"_66":1949,"_1888":27},{"_1877":1928,"_99":1929,"_1851":1930,"_1881":1866,"_1931":1932,"_53":1933,"_14":1934,"_66":1935,"_1888":27},[1936],859.0021,"https://image.mux.com/ZfDsqm02Bjw4U1qiGt2XcKYLXgQo4sdL02/thumbnail.png","poster","mux","2026-10-26T15:25:00.000Z","safety-for-autonomous-organizations","Safety for Autonomous Organizations",{"_203":1937,"_26":205,"_206":1938,"_208":1939,"_210":211},"exsafe0",[1940],[],{"_203":1941,"_26":234,"_235":1942,"_237":1943},"exsafe0s",[],"Axel Backlund, cofounder of Andon Labs and creator of Vending-Bench, shares what agents running vending machines, a market, a cafe, and a radio station reveal about long-horizon autonomy, and why models that can tell when they are in a simulation make real-world deployments essential to safety work.",[1950],1381.3823,"https://image.mux.com/v1FWaHcoYP7EpNyiNbxkyvFqG02JeSzx7/thumbnail.png","2026-10-26T15:30:00.000Z","performance-engineering-for-agents-lessons-from-glean","Performance Engineering for Agents: Lessons from Glean",{"_203":1951,"_26":205,"_206":1952,"_208":1953,"_210":211},"exglean0",[1954],[],{"_203":1955,"_26":234,"_235":1956,"_237":1957},"exglean0s",[],"Tony Gentilcore, cofounder of Glean and a veteran of Chrome's Speed Team, brings a performance-first lens to long-horizon agents in a fireside with Chad Metcalf: reading traces, token budgets that change agent behavior, what work belongs in the background, and measuring spend per unit of throughput.",[1964],1056.5995,"https://image.mux.com/4jMAtMC7uNlI01kBXJ00i6R00SWSPhae005l/thumbnail.png","2026-10-26T15:35:00.000Z","the-mythical-agent-month","The Mythical Agent-Month",{"_203":1965,"_26":205,"_206":1966,"_208":1967,"_210":211},"exwes0",[1968],[],{"_203":1969,"_26":234,"_235":1970,"_237":1971},"exwes0s",[],"Wes McKinney, cofounder of Kenn Software, creator of pandas, and co-creator of Apache Arrow, revisits The Mythical Man-Month for the age of agent factories: why coordination is part of the work, and why judgment and taste matter more as agents take on more of the building.",{"_203":1973,"_26":205,"_206":1974,"_208":1975,"_210":211},"exmem0",[1976],[],{"_203":1977,"_26":234,"_235":1978,"_237":74},"exmem0s",[],{"_103":21,"_1647":1980},"ge7qSe37EQ3jcYkO7utCk8","ua","Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko; compatible; ClaudeBot/1.0; +claudebot@anthropic.com)","hideBanner",{"_2003":2004,"_2005":2006,"_2007":2008,"_2009":2010},"ENV",{"_1987":1988,"_1989":1990,"_1991":1992,"_1993":1994,"_1995":1996,"_1997":1998,"_1999":2000,"_2001":1988,"_2002":1988},"SANITY_DATASET","production","ALGOLIA_INDEX","heavybit-library","ALGOLIA_APPLICATION_ID","0QCKBC1FMK","ALGOLIA_SITEWIDE_INDEX","heavybit-sitewide","ALGOLIA_JOBS_INDEX","heavybit-jobs","MUX_ENV_KEY","dRI98bL8ze03gYDL6-Ogle2Js","ALGOLIA_SEARCH_API_KEY","7733fdd199ec7af7ee2923ec39c3f036","MODE","VITE_CONTEXT","footer",{"_2161":2162,"_2163":2164,"_2165":2166},"globalConfig",{"_75":2158},"header",{"_2047":2048,"_2049":2050},"skills",{"_2011":2012,"_2013":2014,"_2015":2016,"_2017":2018,"_2019":2014,"_2020":2021,"_2022":2023,"_2024":2025,"_2026":2027,"_2028":2029,"_2030":2031,"_2032":2033,"_2034":2035},"cardButtonText","View skill","connectHeading","Connect your agent","connectSteps","Add the endpoint below to your harness: Claude Code, Codex, Claude, Cursor, ChatGPT, or another that supports MCP.\nSign in to your Heavybit account with Google or GitHub when your agent asks. It's free, and your first sign-in creates the account.\nTell your agent what you're working on. It finds the right skill, or you can ask for one by name.","crucibleCta",{"_1543":2036,"_2037":2014,"_1626":2021,"_66":2038},"crucibleLabel","crucibleUrl","https://crucible.heavybit.com/mcp","drawsFromLabel","Built from what these operators shared","explainerBody","A skill is a set of working instructions your agent loads when you need it, built from the strategies operators shared in the Heavybit Library. It guides the work and you still do it. Heavybit reads and approves every skill before it ships, and keeps updating skills as new content enters the Library. Skills are plain text, so you or your agent can read any skill in full before acting on it. Your agent decides how to apply a skill, so check its work; if a skill gets something wrong, ask your agent to report it or email platform@heavybit.com.","explainerHeading","What a skill is","searchPlaceholder","Search skills: pricing, launch, SOC 2...","skillEyebrow","Heavybit Skill","sourcesHeading","Built from the Heavybit Library","sourcesIntro","The operators behind this skill shared what they know in these talks, podcasts, and articles. Every one is free in the Heavybit Library.",[2039],"label","Connect your agent to Crucible, Heavybit's MCP server",{"_203":2040,"_26":205,"_206":2041,"_208":2042,"_210":211},"intro",[2043],[],{"_203":2044,"_26":234,"_235":2045,"_237":2046},"intro0",[],"One endpoint gives your agent every Heavybit Skill, a search of the Heavybit Library, and a direct line to Heavybit. Sign in free with Google or GitHub. Skills load when your agent needs them, so there is nothing to download or update.","announcementBar",{"_2150":2151,"_2152":-5,"_2153":33,"_2154":2155,"_237":2156,"_93":2157},"headerNavLinks",[2051,2052,2053,2054,2055],{"_203":2131,"_2057":2132,"_2037":2133,"_1626":2134,"_2060":12,"_66":2135},{"_203":2116,"_2057":2117,"_2037":2118,"_1626":2119,"_2060":12,"_66":2120},{"_203":2098,"_2057":2099,"_2037":2100,"_1626":2101,"_2060":12,"_66":2102},{"_203":2062,"_2057":2063,"_2037":2064,"_1626":2065,"_2060":12,"_66":2066},{"_203":2056,"_2057":-5,"_2037":2058,"_1626":2059,"_2060":33,"_66":2061},"3702498a502c","dropdown","Contact","/press","mobileOnly","Navigate to contact page","4bec0dd665d4",[2067,2068,2069,2070,2071,2072,2073],"Library","/library","Navigate to library page",{"_203":2095,"_2037":2096,"_1626":2097,"_66":2096},{"_203":2091,"_2037":2092,"_1626":2093,"_66":2094},{"_203":2087,"_2037":2088,"_1626":2089,"_66":2090},{"_203":2083,"_2037":2084,"_1626":2085,"_66":2086},{"_203":2080,"_2037":2081,"_1626":2082,"_66":2081},{"_203":2077,"_2037":2078,"_1626":2079,"_66":2078},{"_203":2074,"_2037":2075,"_1626":2076,"_66":2075},"66773779036c","DevToolsDigest","/devtoolsdigest","2ead3bd14715","Reports","/reports","ac06751f0372","Collections","/library/collections","1f3fb936f3b9","Podcast Network","/library/podcasts","Navigate to podcasts page","6526c69f5260","Videos","/library?filter=videos","Watch educational videos in Heavybit's Library","63e48d609548","Articles","/library?filter=articles","Read educational articles in Heavybit's Library","76b0479bfae2","Skills","/library/skills","716b76469486",[2103,2104,2105],"Community","/community","Navigate to community page",{"_203":2113,"_2037":2114,"_1626":2115,"_66":2114},{"_203":2110,"_2037":2111,"_1626":2112,"_66":2111},{"_203":2106,"_2037":2107,"_1626":2108,"_66":2109},"23077fca0e2f","DevGuild","/devguild","Navigate to devguild page","1fb7c9c465d0","Advisors","/community/advisors","3a177efc4649","Events","/events","90b68b861c06",[2121,2122],"Portfolio","/portfolio","Navigate to portfolio page",{"_203":2127,"_2037":2128,"_1626":2129,"_66":2130},{"_203":2123,"_2037":2124,"_1626":2125,"_66":2126},"e50fa43ad2e7","Jobs","/jobs","Portfolio Company Jobs","d481fcbbd3af","Spotlights","/portfolio/spotlights","Navigate to spotlights page","a8d641babed2",[2136,2137,2138],"About","/about","Navigate to about page",{"_203":2146,"_2037":2147,"_1626":2148,"_66":2149},{"_203":2142,"_2037":2143,"_1626":2144,"_66":2145},{"_203":2139,"_2037":2140,"_1626":2059,"_66":2141},"84e68f9f5c33","News","Navigate to press page","3d279775d73b","FAQ","/faq","Navigate to FAQ page","964627410473","Team","/team","Navigate to team page","buttonText","RSVP","end","show","start","2026-10-20T21:46:59.960Z","🐺 [Oct 28] GitBlame: Werewolf Happy Hour @GitHub Universe","https://luma.com/gitblame-werewolf",{"_26":1851,"_77":2159},{"_1619":2160,"_26":1621},"image-32fdfae4fd13b8f9a3c5bf6b13a635068cb0ecbb-1332x529-png","emailSubscription",{"_1543":2234,"_2235":2236},"footerColumns",[2198,2199,2200],"socialMedia",[2167,2168,2169,2170,2171],{"_203":2193,"_2173":2194,"_93":2195},{"_203":2188,"_2173":2189,"_93":2190},{"_203":2183,"_2173":2184,"_93":2185},{"_203":2178,"_2173":2179,"_93":2180},{"_203":2172,"_2173":2174,"_93":2175},"639729864c2d","icon",{"_26":1617,"_77":2176},"https://www.tiktok.com/@heavybitindustries",{"_1619":2177,"_26":1621},"image-48ac9473e30341c4bee9f6634b77b6f8f1dd6b2f-849x980-webp","a100f6448036",{"_26":1617,"_77":2181},"https://www.instagram.com/heavybitindustries/",{"_1619":2182,"_26":1621},"image-b38b7ee9ccd05f61ca43d5b29fe7058f822f87eb-512x512-webp","fae1cc0df5c6",{"_26":1617,"_77":2186},"https://www.youtube.com/c/heavybit",{"_1619":2187,"_26":1621},"image-6c548d2e815d1fb3b3077c9b8be1ec251e1bc418-512x512-png","5348b028f0d41c3dac62b3dd2b5531ee",{"_26":1617,"_77":2191},"https://www.linkedin.com/company/heavybit",{"_1619":2192,"_26":1621},"image-01bd92eb06873bdc30b1cb5f9c78b1ef5a9559d4-512x512-png","ee025c96397c",{"_26":1617,"_77":2196},"https://twitter.com/heavybit",{"_1619":2197,"_26":1621},"image-9d618da0948299428b29e5d499af3fa5369a24d2-20x20-svg",{"_203":2226,"_2202":2133,"_1870":2227},{"_203":2214,"_2202":2100,"_1870":2215},{"_203":2201,"_2202":2064,"_1870":2203},"ea618aa34610","heading",[2204,2205,2206,2207],{"_203":2211,"_2037":2212,"_1626":2065,"_66":2213},{"_203":2210,"_2037":2084,"_1626":2085,"_66":2086},{"_203":2209,"_2037":2075,"_1626":2076,"_66":2075},{"_203":2208,"_2037":2096,"_1626":2097,"_66":2096},"5656520aaacf","96a97ceba3a2","e1673433290e","c82a64edc4dd","Blog","Navigate to browse on library page","a584a6acf07b",[2216,2217,2218,2219],{"_203":2225,"_2037":2114,"_1626":2115,"_66":2114},{"_203":2224,"_2037":2107,"_1626":2108,"_66":2109},{"_203":2223,"_2037":2124,"_1626":2125,"_66":2126},{"_203":2220,"_2037":2221,"_1626":2222,"_66":2221},"e22bfa6b2330","Inclusion","/community/diversity-and-inclusion","feb038ce46a1","e2d5641ef7de","4358699f4ca4","30a87ef6e7c1",[2228,2229,2230],{"_203":2233,"_2037":2118,"_1626":2119,"_66":2120},{"_203":2232,"_2037":2147,"_1626":2148,"_66":2149},{"_203":2231,"_2037":2140,"_1626":2059,"_66":2141},"398a9b50e214","42666d0b3987","3dfef1283ab5","Unlock success from technical pioneers in software infrastructure. We send the best strategies, tools, and industry updates directly to your inbox.","targetURL","https://share.hsforms.com/1_KYaR4lAQm-C8_9H7r4MIwqdoyu"]

Persistent Memory + What Breaks in Production

Speakers