Meta AI's 'Remember When It Matters' Gives Agents an Active Memory Manager

Source: arXiv — 2026-07-09

Summary

Meta AI researchers identify "behavioral state decay" — the failure mode where decision-relevant context gets buried or evicted as an agent's trajectory grows — and propose a dedicated memory agent that actively decides when to inject reminders into a long-horizon agent's context, rather than passively waiting to be queried. The approach reports +8.3 percentage points on Terminal-Bench 2.0 and +6.8pp on τ²-Bench over baselines without active memory intervention.

Key Takeaways

• Names and targets "behavioral state decay": earlier commitments/constraints getting lost as agent trajectories grow long.
• Proposes a separate memory agent that actively injects reminders at the right moment, instead of a passive retrieve-on-request memory store.
• +8.3pp on Terminal-Bench 2.0 and +6.8pp on τ²-Bench — meaningful gains on long-horizon agentic benchmarks.
• Distinct from retrieval-style memory approaches (e.g. Oracle's hybrid search, Microsoft's Memora) in treating memory injection as an active decision, not passive lookup.

Discussion

(No questions yet — ask follow-ups via a Claude Code chat session on this repo; answers get appended here.)