--- schema: 1 type: video title: 'Agent Memory Is Solved. Agent Learning Isn''t. — Karthik Ranganathan, Yugabyte' description: >- One agent can remember. A team of agents still can't learn together. Karthik Ranganathan, co-founder and co-CEO of Yugabyte, explains why per-agent memory is a solved problem but shared learning across agents isn't. When one agent hands work to another, it passes on the output but loses the reasoning, the dead ends and the context, so the next agent burns tokens re-deriving it all. He covers five misconceptions about memory, what Yugabyte got wrong building its own agents, and how fixing its RAG pipeline one variable at a time took faithfulness from 14% to 82% with about one-seventh the context. Then Heather Downing demos Meko, Yugabyte's agent-native persistence layer, resuming sessions across agents and promoting private memories into shared data packs. In this talk: • Why more memory, shared files and bigger context windows aren't learning • Shared context that's governed, promotable and traceable to its source • How tuning a RAG pipeline cut context size about 7x while raising precision • Demo: resuming and sharing context across Claude, Codex and other agents SPEAKERS Karthik Ranganathan, Co-founder & co-CEO, Yugabyte LinkedIn: https://www.linkedin.com/in/kranganathan X: https://x.com/karthikr Heather Downing, Developer Advocate, Yugabyte LinkedIn: https://www.linkedin.com/in/heathermdowning/ X: https://x.com/quorralyne LINKS Meko: https://mekodata.ai/ Meko docs: https://docs.mekodata.ai/ Meko skills (GitHub): https://github.com/yugabyte/meko-skills Yugabyte: https://www.yugabyte.com/ CHAPTERS 0:00 Intro 0:12 Memory is solved, learning isn't 0:47 Learning happens in groups 1:17 Why a database company built this 2:02 A year of failing to ship agents 2:37 The real failure: lost context 3:27 The new-teammate problem 3:57 Five misconceptions 5:02 Shared, governed, traceable context 5:42 What we got wrong 6:37 Fixing RAG: 14% → 82% faithfulness 7:27 Introducing Meko and data packs 8:36 This deck was built with Meko 10:20 Demo: collective knowledge 11:25 Sharing context saves tokens 12:20 Demo: saving to a data pack 13:25 Demo: resuming in a fresh session 14:20 Demo: sharing with another agent 14:55 Private vs shared memory 16:15 Proving the token savings 16:45 Human promotion as a training signal 17:52 How the team uses Meko 19:26 Capturing tribal knowledge 19:40 Try Meko Recorded at the AI Engineer World's Fair 2026 in San Francisco. Subscribe for more talks from the engineers building with AI. AI Engineer: https://ai.engineer YouTube: https://www.youtube.com/@aiDotEngineer X: https://x.com/aiDotEngineer LinkedIn: https://www.linkedin.com/company/aidotengineer/ #AgentMemory #AIAgents #AIEngineer canonical: 'https://liquidgold.cc/watch/9_oy4owGtPI' markdown: 'https://liquidgold.cc/watch/9_oy4owGtPI.md' llms: 'https://liquidgold.cc/watch/9_oy4owGtPI/llms.txt' parent: 'https://liquidgold.cc/channel/aidotengineer.md' channel: aidotengineer updated: '2026-10-23T22:19:59.028Z' language: null children: [] videoId: 9_oy4owGtPI durationSeconds: 1236 publishedAt: '2026-10-05T14:00:20.000Z' searchChunks: - ordinal: 0 text: null startSeconds: 24 endSeconds: 56 watch: 'https://liquidgold.cc/watch/9_oy4owGtPI?t=24' heldReason: rights - ordinal: 1 text: null startSeconds: 56 endSeconds: 68 watch: 'https://liquidgold.cc/watch/9_oy4owGtPI?t=56' heldReason: rights - ordinal: 2 text: null startSeconds: 171 endSeconds: 214 watch: 'https://liquidgold.cc/watch/9_oy4owGtPI?t=171' heldReason: rights - ordinal: 3 text: null startSeconds: 214 endSeconds: 225 watch: 'https://liquidgold.cc/watch/9_oy4owGtPI?t=214' heldReason: rights - ordinal: 4 text: null startSeconds: 246 endSeconds: 256 watch: 'https://liquidgold.cc/watch/9_oy4owGtPI?t=246' heldReason: rights - ordinal: 5 text: null startSeconds: 256 endSeconds: 310 watch: 'https://liquidgold.cc/watch/9_oy4owGtPI?t=256' heldReason: rights - ordinal: 6 text: null startSeconds: 310 endSeconds: 320 watch: 'https://liquidgold.cc/watch/9_oy4owGtPI?t=310' heldReason: rights - ordinal: 7 text: null startSeconds: 320 endSeconds: 352 watch: 'https://liquidgold.cc/watch/9_oy4owGtPI?t=320' heldReason: rights - ordinal: 8 text: null startSeconds: 363 endSeconds: 397 watch: 'https://liquidgold.cc/watch/9_oy4owGtPI?t=363' heldReason: rights - ordinal: 9 text: null startSeconds: 407 endSeconds: 454 watch: 'https://liquidgold.cc/watch/9_oy4owGtPI?t=407' heldReason: rights - ordinal: 10 text: null startSeconds: 407 endSeconds: 454 watch: 'https://liquidgold.cc/watch/9_oy4owGtPI?t=407' heldReason: rights - ordinal: 11 text: null startSeconds: 464 endSeconds: 521 watch: 'https://liquidgold.cc/watch/9_oy4owGtPI?t=464' heldReason: rights - ordinal: 12 text: null startSeconds: 631 endSeconds: 685 watch: 'https://liquidgold.cc/watch/9_oy4owGtPI?t=631' heldReason: rights - ordinal: 13 text: null startSeconds: 697 endSeconds: 741 watch: 'https://liquidgold.cc/watch/9_oy4owGtPI?t=697' heldReason: rights - ordinal: 14 text: null startSeconds: 751 endSeconds: 809 watch: 'https://liquidgold.cc/watch/9_oy4owGtPI?t=751' heldReason: rights - ordinal: 15 text: null startSeconds: 896 endSeconds: 908 watch: 'https://liquidgold.cc/watch/9_oy4owGtPI?t=896' heldReason: rights - ordinal: 16 text: null startSeconds: 908 endSeconds: 985 watch: 'https://liquidgold.cc/watch/9_oy4owGtPI?t=908' heldReason: rights - ordinal: 17 text: null startSeconds: 985 endSeconds: 996 watch: 'https://liquidgold.cc/watch/9_oy4owGtPI?t=985' heldReason: rights - ordinal: 18 text: null startSeconds: 1008 endSeconds: 1019 watch: 'https://liquidgold.cc/watch/9_oy4owGtPI?t=1008' heldReason: rights - ordinal: 19 text: null startSeconds: 1084 endSeconds: 1173 watch: 'https://liquidgold.cc/watch/9_oy4owGtPI?t=1084' heldReason: rights --- # Agent Memory Is Solved. Agent Learning Isn't. — Karthik Ranganathan, Yugabyte One agent can remember. A team of agents still can't learn together. Karthik Ranganathan, co-founder and co-CEO of Yugabyte, explains why per-agent memory is a solved problem but shared learning across agents isn't. When one agent hands work to another, it passes on the output but loses the reasoning, the dead ends and the context, so the next agent burns tokens re-deriving it all. He covers five misconceptions about memory, what Yugabyte got wrong building its own agents, and how fixing its RAG pipeline one variable at a time took faithfulness from 14% to 82% with about one-seventh the context. Then Heather Downing demos Meko, Yugabyte's agent-native persistence layer, resuming sessions across agents and promoting private memories into shared data packs. In this talk: • Why more memory, shared files and bigger context windows aren't learning • Shared context that's governed, promotable and traceable to its source • How tuning a RAG pipeline cut context size about 7x while raising precision • Demo: resuming and sharing context across Claude, Codex and other agents SPEAKERS Karthik Ranganathan, Co-founder & co-CEO, Yugabyte LinkedIn: https://www.linkedin.com/in/kranganathan X: https://x.com/karthikr Heather Downing, Developer Advocate, Yugabyte LinkedIn: https://www.linkedin.com/in/heathermdowning/ X: https://x.com/quorralyne LINKS Meko: https://mekodata.ai/ Meko docs: https://docs.mekodata.ai/ Meko skills (GitHub): https://github.com/yugabyte/meko-skills Yugabyte: https://www.yugabyte.com/ CHAPTERS 0:00 Intro 0:12 Memory is solved, learning isn't 0:47 Learning happens in groups 1:17 Why a database company built this 2:02 A year of failing to ship agents 2:37 The real failure: lost context 3:27 The new-teammate problem 3:57 Five misconceptions 5:02 Shared, governed, traceable context 5:42 What we got wrong 6:37 Fixing RAG: 14% → 82% faithfulness 7:27 Introducing Meko and data packs 8:36 This deck was built with Meko 10:20 Demo: collective knowledge 11:25 Sharing context saves tokens 12:20 Demo: saving to a data pack 13:25 Demo: resuming in a fresh session 14:20 Demo: sharing with another agent 14:55 Private vs shared memory 16:15 Proving the token savings 16:45 Human promotion as a training signal 17:52 How the team uses Meko 19:26 Capturing tribal knowledge 19:40 Try Meko Recorded at the AI Engineer World's Fair 2026 in San Francisco. Subscribe for more talks from the engineers building with AI. AI Engineer: https://ai.engineer YouTube: https://www.youtube.com/@aiDotEngineer X: https://x.com/aiDotEngineer LinkedIn: https://www.linkedin.com/company/aidotengineer/ #AgentMemory #AIAgents #AIEngineer Canonical: https://liquidgold.cc/watch/9_oy4owGtPI · Channel: [AI Engineer](https://liquidgold.cc/channel/aidotengineer.md) > Search chunk text is held until public-source provenance is verified. ## Document chunks --- #### Chunk 0: Document excerpt (0:13 to 0:56) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 1: Document excerpt (0:56 to 1:19) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 2: Document excerpt (1:19 to 2:03) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 3: Document excerpt (2:03 to 2:38) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 4: Document excerpt (2:38 to 3:34) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 5: Document excerpt (3:34 to 4:06) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 6: Document excerpt (4:06 to 5:10) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 7: Document excerpt (5:10 to 5:52) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 8: Document excerpt (5:52 to 6:37) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 9: Document excerpt (6:37 to 7:34) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 10: Document excerpt (7:34 to 8:41) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 11: Document excerpt (8:41 to 10:20) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 12: Document excerpt (10:20 to 11:25) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 13: Document excerpt (11:25 to 12:21) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 14: Document excerpt (12:21 to 13:29) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 15: Document excerpt (13:29 to 14:24) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 16: Document excerpt (14:24 to 15:08) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 17: Document excerpt (15:08 to 16:48) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 18: Document excerpt (16:48 to 17:52) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 19: Document excerpt (17:52 to 19:33) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 20: Document excerpt (19:33 to 20:36) > Held: transcript rights do not permit this chunk's text.