Meta gave its AI agent a second agent whose only job is remembering things
Agents on long tasks forget constraints and repeat failed commands. Meta AI's answer is a separate memory agent that keeps a structured record and decides when to interrupt with a reminder, worth up to 8 points on two benchmarks.
Anyone who has watched an AI agent work through a long task has seen the pattern. It agrees to a constraint early on, then quietly breaks it an hour later while fixing something unrelated. It runs a command, gets an error, and tries a nearly identical command a few minutes later. It diagnoses a problem, then meets the same problem again and treats it as brand new. Meta AI researchers have given this a name, behavioral state decay, and published a system for dealing with it.
Their diagnosis is that the information guiding the agent’s decisions gets scattered through a growing task history. It ends up buried deep in the context window, which is roughly how much text the model can hold in mind at once, or it falls out entirely. Crucially, Meta says simply giving agents a longer memory does not fix this. Even when the relevant line is still technically in the transcript, it stops shaping what the agent does.
The proposed fix pairs an ordinary action agent with a separate memory agent that runs alongside it. Every few steps, the memory agent reviews the recent history, updates a structured memory bank, and then makes one decision: add a short reminder to the action agent’s next call, or stay quiet. The staying quiet part matters, because too many reminders cost tokens, add delay and distract from the work at hand.
The memory bank has three parts. A private status field tracking progress and open risks, which the action agent never sees. A knowledge section for stable facts like file paths and requirements. And a procedural section recording what was tried and what happened, including failed commands and rejected hypotheses.
The results are modest but consistent. With Claude Sonnet 4.5 as the action agent, the system solved 46 percent of Terminal-Bench 2.0 tasks on the first attempt against a baseline of 38 percent. On tau2-Bench, which tests tool use in airline, retail and telecom scenarios, the average rose from 55 to 62 percent. Gains were larger for the weaker model but did not vanish with a stronger one. A nice example from the airline tests: a user claimed Gold status while the system’s own tool said otherwise. The baseline granted compensation based on the claim. The memory agent reminded the action agent to trust the verified data instead.
What is behind this. The industry spent two years treating context length as the answer to agent forgetfulness, and this paper is part of a growing argument that it was the wrong lever. Attention is not the same as storage. What Meta is really proposing is an editor sitting beside the writer, deciding what deserves to be said again. Ablation tests support that reading: feeding the agent the full memory bank at every step performed worse than the selective version, and worse than doing nothing in some domains.
What this means for you: for most people, nothing today. This is plumbing, and it will show up inside tools rather than as a feature you switch on. If you build with agents, the code is on GitHub and the design is meant to be plug and play with existing harnesses. And if you simply use agents for long jobs, the useful habit is the human version of the same idea: restate your key constraints partway through a long session rather than assuming the model still has them in view.
Sources
Source: https://arxiv.org/abs/2607.08716
Researchers gave an AI agent a real company, a bank card and 24 hours. It lost 447 dollars
Bottleneck Labs handed GPT-5.6 Sol a live iOS app, a Mac mini, 350 dollars and one instruction: grow the business. The agent coded well, then bought fake users, spammed its own customers and cut the price six times.
