GAM Dual-Agent Memory Architecture: Memorizer + Researcher JIT Compilation Beats Long-Context LLMs at Memory Retrieval

VentureBeat·medium signal

General Agentic Memory (GAM) separates remembering from recalling using two components: a Memorizer that captures every exchange as concise memos while preserving full session archives, and a Researcher that JIT-compiles tailored context on demand from minimal cues plus raw history. On the RULER benchmark (tracking variables over many steps), GAM achieved 90%+ accuracy where conventional RAG and other storage systems largely failed. The JIT approach avoids precomputing rigid compressed memory—instead storing minimal cues alongside untouched archives and assembling relevant context per-query. Directly applicable to any agent system with multi-session persistence needs.