Meta AI Launches Memory Coach Agent: Accuracy Up 8.3 Percentage Points

Meta AI has published a description of a two-agent system in which a helper coach agent maintains a structured memory bank and decides when to remind the main agent of errors it has already diagnosed. On two benchmarks, the system improved results by up to 8.3 percentage points.

Why agents keep repeating the same mistakes

AI agents handle short instructions well, but on long multi-step tasks they begin to "forget" what they have already figured out. An agent may diagnose a problem at step five, then try the same failed path again a few iterations later. The details have slipped out of its current attention window, and the system falls into a loop.

This is not an abstract problem. In a typical scenario, an agent finds that a certain approach does not work and adjusts its strategy. Over the course of a long task, though, it "forgets" this and returns to the original broken step. Repeated error cycles are one of the main reasons agents stall on real-world tasks such as writing code, research, and multi-stage automation.

How the memory coach agent works

Meta AI addressed the problem with a separate, specialized role: the memory coach agent. It does not perform the main task. Instead, it maintains a structured data bank of what the main agent has tried and what it has found.

The helper agent records the following in the memory bank:

• errors that have already been diagnosed, and their causes
• failed steps that should be abandoned
• patterns that have not produced results before
• current context about the state of the task

The key feature of the system is selective intervention. The coach agent does not interrupt the main agent at every step. It decides for itself: if the situation calls for a reminder, it steps in; if there is no risk, it stays silent. Knowing when to stay quiet turned out to be just as important as the reminder itself, because constant interruptions would overload the main agent with extra context and reduce its effectiveness.

"A specialized coach agent maintains a structured memory bank and decides when to remind the main agent and when to stay silent," according to

Meta AI's description of the system.

How much did benchmark quality improve?

On two benchmarks, the system with the coach agent showed a gain of up to 8.3 percentage points compared with a single agent without a memory-management mechanism. According to Meta AI, the tests covered complex multi-step tasks, the scenarios in which single agents most often lose relevant context and start looping.

A gain of just over 8 percentage points sounds modest, but for agentic tasks, where baseline accuracy is often 40–60%, it is a significant step forward. Meta AI's published announcement does not name the specific benchmarks used.

What this means

The "main agent + memory coach" architecture is a practical answer to one of the key flaws of current AI systems. The problem is not a lack of "intelligence" in the agent, which is capable of diagnosing errors correctly. The problem is the lack of a structured mechanism that keeps those diagnoses in working context throughout a long session.

Meta AI moves memory management into a separate, specialized role and shows a measurable result. If the approach proves scalable, it could make agents more reliable on real long-running tasks: code automation, research pipelines, and multi-stage business processes.

*Meta is designated an extremist organization and banned in Russia.