New system pairs a dedicated memory agent with existing language models to combat 'behavioral state decay' during extended tasks Meta AI researchers have built what amounts to a personal assistant for AI personal assistants. Their new Proactive Memory Agent, or PMA, sits alongside a standard large language model and whispers relevant reminders at just the right moments, boosting performance on complex, multi-step tasks without touching the underlying model itself.
The research, published in the paper “Remember When It Matters: Proactive Memory Agent for Long-Horizon Agents” (arXiv:2607.08716) in July 2026, tackles a problem anyone who’s used an AI agent for extended work sessions has probably noticed: the longer a task runs, the worse the AI gets at remembering what it’s supposed to be doing.
The forgetting problem, solved with a sidekick #
Meta’s researchers call this degradation “behavioral state decay.” The PMA architecture pairs an unmodified action agent with a separate memory agent that maintains a structured memory bank organized into three categories: knowledge sections that store stable facts and requirements, procedural sections that record previous attempts, failures, and fixes, and private status sections.
The memory agent monitors a sliding window of the action agent’s recent steps. When it detects the action agent drifting off course or about to repeat a known mistake, it injects a concise reminder into the agent’s context. When things are going fine, it stays quiet.
Traditional memory systems in AI generally fall into two camps: passive retrieval, where memory is only accessed when explicitly queried, and always-active reminders, where context is constantly stuffed with potentially relevant information. The PMA outperformed both approaches in Meta’s experiments.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
The numbers behind the nudges #
When researchers used Claude Sonnet 4.5 as the action agent with the PMA system, pass@1 success rates on Terminal-Bench 2.0 jumped from 37.6% to 45.9%, an increase of 8.3 percentage points. On the τ²-Bench benchmark, rates climbed from 55.0% to 61.8%, a gain of 6.8 percentage points.
Claude Opus 4.6, already a more capable model, still saw gains of 2.4 percentage points on Terminal-Bench 2.0 and 2.5 percentage points on τ²-Bench when paired with the PMA. The best results came when Claude Opus 4.6 served as the memory agent itself.
From research paper to consumer product #
A consumer product called Muse launched on September 8, 2026, integrating proactive memory features into a personal AI assistant designed for task execution with long-term personalization capabilities. The product aims to bring persistent memory and goal tracking to everyday users.
Because the memory agent works alongside an unmodified action model, it could theoretically be paired with any sufficiently capable LLM. This modular design means developers building AI agents don’t need to wait for the next generation of foundation models to get meaningfully better performance on complex, multi-step workflows.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our