Memory is the Agent: Architecting Stateful Reasoning
summary
Shows a concrete memory layer (Python + Postgres demo on GitHub) that flips a stateless LLM into a stateful agent by read/write memory per turn; A/B demo toggles memory OFF → ON so the same model and tools produce different plans. Memory items use Identity, Principles, Focus, Signals plus fields: type, source, timestamp, priority, TTL (60 days), append‑only versioning for rollbacks. Retrieval uses a hybrid index with semantic similarity, recency decay and priority (relevance 60% / recency 25% / priority 15%), returning top 10 with P50 ~150ms, P95 ~400ms.
AI-written summary · about
chapters (21)
- 0:00Introduction and speaker intro
- 0:23Core idea of memory as agent
- 1:48Evolution from chatbots to agents
- 3:12Memory as stateful reasoning
- 4:15Demo of journaling bot
- 5:23Continuity in agent behavior
- 6:06Demo continuity with memory off
- 7:56Demo continuity with memory on
- 8:19Latency and retrieval challenges
- 9:22Principles of memory design
- 10:04Memory item structure
- 12:12Retrieval and ranking factors
- 14:02Teachability and correction
- 14:45Demo teachability edit
- 15:51Versioning for safety
- 17:19Learning and evolution
- 18:47Demo learning over time
- 20:14Pattern recognition in memory
- 21:37Three superpowers combined
- 21:58Key takeaways and design principles
- 23:24Closing and resources
