Meta AI запустила агента-наставника памяти: точность выросла на 8,3 п.п.

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

Why agents repeat the same mistakes over and over

AI agents handle short instructions well, but on long multi-step tasks they start to "forget" what they have already discovered. An agent may diagnose a problem at the fifth step, then after several iterations attempt the same failing path again — the details have moved beyond the horizon of its current attention, and the system enters a loop.

This is not an abstract problem. A typical scenario: an agent discovers that a certain approach does not work, adjusts its strategy — but during a long task "forgets" about this and returns to the original non-working step. Repeated error cycles are one of the main reasons agents stall on real-world tasks: writing code, research, multi-step automations.

How the memory mentor agent works

Meta AI solved this problem with a separate specialized role — the memory mentor agent. It does not perform the main task; instead, it maintains a structured database of what the main agent has tried and discovered.

The auxiliary agent records in the memory bank:

• already-diagnosed errors and their causes
• failed steps that should be abandoned
• patterns that previously yielded no results
• current context about the state of the task

The key feature of the system is the selectivity of its interventions. The mentor agent does not interrupt the main agent at every step. It makes its own decision: if the situation requires a reminder — intervene; if there is no risk — remain silent. The ability to "stay silent at the right moment" turned out to be no less important than the reminder itself: constant interruptions would overload the main agent with unnecessary context and reduce its effectiveness.

"A specialized mentor agent maintains a structured memory bank and decides when to remind the main agent and when to remain silent," the

Meta AI system description states.

How much quality improved on tests

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

A gain of over 8 percentage points sounds modest, but for agentic tasks where the baseline accuracy is often 40–60%, this is a significant step forward. The specific names of the benchmarks used are not provided in Meta AI's published announcement.

What this means

The "main agent + memory trainer" architecture is a practical response to one of the key flaws in modern AI systems. The problem is not a lack of "intelligence" in the agent: it is capable of correctly diagnosing errors. The problem is the absence of a structured mechanism that would retain these diagnoses in the working context throughout a long session.

Meta AI moves memory management into a separate specialized role — and demonstrates a measurable result. If the approach proves scalable, it could increase the reliability of agents in real-world long-term tasks: code automation, research pipelines, multi-step business processes.

*Meta is recognized as an extremist organization and is banned in Russia.