The Anatomy of Memory in Humans and AI Agents
summary
A concrete memory stack: the Radius agent memory server stores session/working memory (namespace, session ID, user ID, messages, TTL) and a long-term store with text, entities, embeddings and access count; it runs background extraction and automatic summarization, supports four extraction strategies (discrete, summary, preferences, custom), and integrates via SDKs or Spring AI advisors. Demo shows extracted facts (Marcela, Pongo) persist across sessions. Notes tradeoffs: attention is quadratic so context accuracy degrades with length, LLMs need retraining to update parameters, and fine‑tuning risks catastrophic forgetting.
AI-written summary · about
chapters (11)
- 0:00Introduction to memory talk
- 1:30Why human memory is complex
- 5:45Atkinson-Shiffrin memory model
- 8:39Squire’s long-term memory types
- 16:54Working memory components
- 20:46LLM parameter vs activation memory
- 28:35Human vs AI memory comparison
- 35:02Radius agent memory server
- 41:22Java Spring AI integration demo
- 50:37Enhancing agent memory strategies
- 52:07Future of AI and neuroscience
related talks
46mThe Anatomy of Memory in Humans and AI Agents
01h 49mBuilding Memory Augmented Agents with MemoRizz
25mWhat does it take to build long-term state for AI agents?
18mArchitecting Agent Memory: Principles, Patterns, and Best Practices
15mMemory-Augmented AI Agents: Enhancing Long-Term Reasoning and Context Awareness
47mAgents and Memory Architecture
28mAgentic Memory
51mSecrets of Agentic Memory Management