What does it take to build long-term state for AI agents?
summary
Panel claims a production memory layer must be LLM-agnostic and handle real-world needs: ingest agent traces, enforce memory pools and TTL with auth, detect supersession handling (new facts replacing old), and run an intelligent recall pipeline that blends nearest-neighbor + keyword ranking to surface ~50 relevant memories from billions. Practical wins cited: CPU-only embedding stacks that cut search costs (claimed ~41% code-bill saving) and LocoMo benchmark results showing a 95.03% reduction in inference spend using memory vs full-context prompts.
AI-written summary · about
chapters (14)
- 0:00Agent memory systems intro
- 1:30Defining agent memory
- 2:53Memory vs context in agents
- 4:17Where semantic retrieval excels and fails
- 5:41Cosine similarity limitations
- 6:46Agent memory ownership
- 8:30What agents should remember/forget
- 10:57Handling non-NL traces
- 12:22Validating autonomous memory
- 14:27Avoiding data snowballing
- 17:37Reducing agentic inference costs
- 20:47Model-agnostic memory
- 22:31Avoiding vendor lock-in
- 24:14Enterprise agent memory mistakes
related talks
46mThe Anatomy of Memory in Humans and AI Agents
15mMemory-Augmented AI Agents: Enhancing Long-Term Reasoning and Context Awareness
27mMemory Engineering: Going Beyond Context Engineering
54mThe Anatomy of Memory in Humans and AI Agents
14mMake Agents Remember the Right Context at the Right Time
47mAgents and Memory Architecture
01h 49mBuilding Memory Augmented Agents with MemoRizz
32mThe AI Operating System: Stateful Agents with Letta