--- schema: 1 type: video title: 'Giving AI Agents Memory That Learns — Jake Broekhuizen, LangChain' description: >- Most teams are great at watching what their agents did. Far fewer turn that into lessons for the next run. Jake Broekhuizen from LangChain's Labs team explains how to give agents memory that actually changes their behavior. Using a real financial services agent whose tone kept slipping from advisory to pushy, he shows why traces and logs are only evidence: they become memory only when a lesson turns into durable context the agent reads next time. He breaks memory into semantic, episodic and procedural types, separates working memory from long-term memory, and walks through the read-write cycle: read context, run, filter the evidence for signal, and write it back. He shows how LangSmith and Context Hub capture, analyze and update that context, and shares three hard-won lessons about what to keep, caching, and when a human should review changes. In this talk: • Why observability alone doesn't make an agent improve • Semantic, episodic and procedural memory, and which one moves behavior most • The read-write cycle: run, filter evidence for signal, update context • Lessons: most traces stay history, watch caching, review procedural changes SPEAKER Jake Broekhuizen, LangChain Labs LinkedIn: https://www.linkedin.com/in/jake-broekhuizen/ LINKS LangChain Labs: https://www.langchain.com/blog/introducing-langchain-labs LangChain on X: https://x.com/LangChain CHAPTERS 0:00 Intro 0:37 Agents that improve every run 1:02 Example: a finance agent's tone slips 2:42 Observing isn't learning 3:07 Traces are everywhere 4:02 From logs to signal 4:22 What memory really is 4:47 Semantic, episodic and procedural memory 5:42 Procedural memory drives behavior 6:22 Working vs long-term memory 8:12 The read-write cycle 9:27 Deciding what's worth keeping 10:22 Capture, analyze, update 10:52 LangSmith and Context Hub 12:02 The finance agent, fixed 13:17 Lesson 1: most traces stay history 13:42 Lesson 2: caching and the hot path 14:47 Lesson 3: human review for key rules 15:27 Wrap-up 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 #LangChain #AIEngineer canonical: 'https://liquidgold.cc/watch/KGFyOtl5ktI' markdown: 'https://liquidgold.cc/watch/KGFyOtl5ktI.md' llms: 'https://liquidgold.cc/watch/KGFyOtl5ktI/llms.txt' parent: 'https://liquidgold.cc/channel/aidotengineer.md' channel: aidotengineer updated: '2026-10-23T22:19:34.240Z' language: null children: [] videoId: KGFyOtl5ktI durationSeconds: 999 publishedAt: '2026-10-08T14:00:37.000Z' searchChunks: - ordinal: 0 text: null startSeconds: 79 endSeconds: 166 watch: 'https://liquidgold.cc/watch/KGFyOtl5ktI?t=79' heldReason: rights - ordinal: 1 text: null startSeconds: 79 endSeconds: 166 watch: 'https://liquidgold.cc/watch/KGFyOtl5ktI?t=79' heldReason: rights - ordinal: 2 text: null startSeconds: 202 endSeconds: 247 watch: 'https://liquidgold.cc/watch/KGFyOtl5ktI?t=202' heldReason: rights - ordinal: 3 text: null startSeconds: 304 endSeconds: 346 watch: 'https://liquidgold.cc/watch/KGFyOtl5ktI?t=304' heldReason: rights - ordinal: 4 text: null startSeconds: 391 endSeconds: 403 watch: 'https://liquidgold.cc/watch/KGFyOtl5ktI?t=391' heldReason: rights - ordinal: 5 text: null startSeconds: 505 endSeconds: 516 watch: 'https://liquidgold.cc/watch/KGFyOtl5ktI?t=505' heldReason: rights - ordinal: 6 text: null startSeconds: 516 endSeconds: 572 watch: 'https://liquidgold.cc/watch/KGFyOtl5ktI?t=516' heldReason: rights - ordinal: 7 text: null startSeconds: 584 endSeconds: 630 watch: 'https://liquidgold.cc/watch/KGFyOtl5ktI?t=584' heldReason: rights - ordinal: 8 text: null startSeconds: 675 endSeconds: 729 watch: 'https://liquidgold.cc/watch/KGFyOtl5ktI?t=675' heldReason: rights - ordinal: 9 text: null startSeconds: 840 endSeconds: 894 watch: 'https://liquidgold.cc/watch/KGFyOtl5ktI?t=840' heldReason: rights - ordinal: 10 text: null startSeconds: 894 endSeconds: 905 watch: 'https://liquidgold.cc/watch/KGFyOtl5ktI?t=894' heldReason: rights --- # Giving AI Agents Memory That Learns — Jake Broekhuizen, LangChain Most teams are great at watching what their agents did. Far fewer turn that into lessons for the next run. Jake Broekhuizen from LangChain's Labs team explains how to give agents memory that actually changes their behavior. Using a real financial services agent whose tone kept slipping from advisory to pushy, he shows why traces and logs are only evidence: they become memory only when a lesson turns into durable context the agent reads next time. He breaks memory into semantic, episodic and procedural types, separates working memory from long-term memory, and walks through the read-write cycle: read context, run, filter the evidence for signal, and write it back. He shows how LangSmith and Context Hub capture, analyze and update that context, and shares three hard-won lessons about what to keep, caching, and when a human should review changes. In this talk: • Why observability alone doesn't make an agent improve • Semantic, episodic and procedural memory, and which one moves behavior most • The read-write cycle: run, filter evidence for signal, update context • Lessons: most traces stay history, watch caching, review procedural changes SPEAKER Jake Broekhuizen, LangChain Labs LinkedIn: https://www.linkedin.com/in/jake-broekhuizen/ LINKS LangChain Labs: https://www.langchain.com/blog/introducing-langchain-labs LangChain on X: https://x.com/LangChain CHAPTERS 0:00 Intro 0:37 Agents that improve every run 1:02 Example: a finance agent's tone slips 2:42 Observing isn't learning 3:07 Traces are everywhere 4:02 From logs to signal 4:22 What memory really is 4:47 Semantic, episodic and procedural memory 5:42 Procedural memory drives behavior 6:22 Working vs long-term memory 8:12 The read-write cycle 9:27 Deciding what's worth keeping 10:22 Capture, analyze, update 10:52 LangSmith and Context Hub 12:02 The finance agent, fixed 13:17 Lesson 1: most traces stay history 13:42 Lesson 2: caching and the hot path 14:47 Lesson 3: human review for key rules 15:27 Wrap-up 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 #LangChain #AIEngineer Canonical: https://liquidgold.cc/watch/KGFyOtl5ktI · 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 (0:12 to 1:08) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 1: Document excerpt (1:08 to 2:46) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 2: Document excerpt (2:46 to 4:07) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 3: Document excerpt (4:07 to 4:52) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 4: Document excerpt (4:52 to 6:31) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 5: Document excerpt (6:31 to 8:25) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 6: Document excerpt (8:25 to 9:32) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 7: Document excerpt (9:32 to 10:30) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 8: Document excerpt (10:30 to 12:09) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 9: Document excerpt (12:09 to 13:26) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 10: Document excerpt (13:26 to 13:48) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 11: Document excerpt (13:48 to 14:54) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 12: Document excerpt (14:54 to 15:28) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 13: Document excerpt (15:28 to 16:39) > Held: transcript rights do not permit this chunk's text.