Make Agents Remember the Right Context at the Right Time
summary
A game called Horcrux Hunt exposes memory failure modes and drives a concrete fix: a three-layer memory model (working, retrieval, persistent) plus off-model belief maps and entropy-based routing that decide when to call an LLM. The implementation compresses state (97% token compression), cuts latency ~75%, reduces cost >80% and doubles win rate in the same agents, demonstrating that structured state, selective retrieval and persistence beat bloated context windows for practical agent reliability.
AI-written summary · about
chapters (12)
- 0:00Why My Agent Failed
- 1:22Horcrux Hunt Setup
- 2:27Memory Failure Modes
- 3:37Context Window Tax
- 4:33Entropy As Memory Lens
- 5:40Three Layer Memory Model
- 6:53Working Memory Compression
- 7:42Structured Belief Maps
- 8:43Entropy Based Routing
- 10:04Persistent Memory Pattern
- 10:55Architecture And Results
- 12:00Five Principles And Wrap
related talks
25mWhat does it take to build long-term state for AI agents?
15mMemory-Augmented AI Agents: Enhancing Long-Term Reasoning and Context Awareness
47mAgents and Memory Architecture
46mThe Anatomy of Memory in Humans and AI Agents
16mHow we solved Context Management in Agents
18mArchitecting Agent Memory: Principles, Patterns, and Best Practices
27mMemory Engineering: Going Beyond Context Engineering
30mDesigning Fast & Cost-Efficient Agents: The Power of Caching