After following a link from my newsreader to a Github repo for General Agentic Memory a few days ago, I found my way to the paper that inspired it and was rewarded with a nifty algorithm for implementic memory for agents. Across the LoCoMo, HotpotQA, and RULER benchmarks, “GAM” beats every memory-based baseline (Mem0, A-Mem, MemoryOS, LightMem) by a healthy margin.
The paper states the reasoning behind the mechanism clearly enough: “Memorization is a form of data compression; thus, it is inevitably subject to information loss.” That’s an easy-enough argument to understand. I stipulate that it’s obviously true.
What they’ve done architecturally is a two-agent split. A Memorizer watches the incoming session stream and writes a lightweight memo for each session that refers to the complete session content that’s dumped in a page-store. Then, a Researcher handles the actual query when it’s time to find some context, first planning a search, then retrieving from the page-store using a mix of dense vector search, BM25, and direct ID lookup, and finally reflecting on whether it’s actually answered the question before deciding to stop or dig again. Both agents get optimized jointly with reinforcement learning against the client’s actual task reward.
I suspect the dual agent and retrieval-reflection approach might be adaptable to other, less agent-memory-specific tasks. It feels like a generalizable pattern.