A research team has built a memory gatekeeper for AI agents that decides whether a new experience is worth remembering — or just noise that will crowd out something more useful.
Large language model agents increasingly rely on sequentially updated memory to reuse past experience. The problem: current systems accept every locally generated update without asking whether it actually helps. That leads to three failure modes — useful knowledge gets overwritten, overly specific rules creep in, and recent examples bias the whole memory store. The proposed fix, called Janus, is a plug-in controller that sits between the memory updater and the memory store. It uses a "Memory Momentum Trigger" to flag suspicious shifts in the update trajectory, then runs old and new memories against a compact hybrid evaluation set — coverage tasks, boundary tasks, and fresh tasks — rather than replaying the entire history. Across six datasets, two backbone LLMs, and two different memory updaters, Janus lifted average accuracy by 2.7 to 4.6 percentage points over the base updaters it wrapped.
The gains matter because Janus is method-agnostic: it wraps existing updaters without touching their internal rules, meaning it could theoretically slot into most agent memory pipelines with minimal friction. That is a more tractable path to better agent memory than redesigning updaters from scratch, and it addresses a failure mode — agents confidently "learning" the wrong lesson — that is easy to overlook when benchmarks only measure final performance.
The 2.7-to-4.6-point accuracy band is modest but consistent, which is more credible than a single headline number. The real test will come when someone deploys this outside controlled datasets, where memory trajectories are messier and evaluation sets harder to define.