AI agents can record a failed command or a critical requirement and still repeat the same mistake later. Meta researchers are proposing a second agent to address that problem: a memory coach that monitors task progress, maintains structured records, and selectively reminds the primary agent what matters.

The approach, described in a new research paper and reported by The Decoder, is designed for long-running tasks in command-line environments and tool-using customer-service scenarios. Meta’s reported tests showed gains of up to 8.3 percentage points across two benchmarks, but the results also point to unresolved questions about intervention frequency, model calibration, operating cost, and how much autonomy the memory system should have.

A memory agent that can choose silence

The researchers call the underlying failure “behavioral state decay.” As an agent’s task history grows, constraints, diagnoses, and previous attempts can become buried in the context window. The information may still be technically available, but it no longer reliably influences the next decision.

Meta’s system separates this problem from conventional memory storage. An unmodified action agent performs the task, while a separate memory agent reviews recent steps at fixed intervals. The memory agent updates a structured memory bank and decides whether the next call to the action agent needs a short reminder—or no reminder at all.

That ability to remain silent is central to the design. Constantly sending the full memory bank back to the action agent can increase token use, latency, and distraction. Meta’s system instead treats intervention as a decision in its own right.

The memory bank has three components. A private status field tracks progress and unresolved risks but is not shown to the action agent. Knowledge Memory stores relatively stable information, such as requirements, file paths, and configurations. Procedural Memory records actions and outcomes, including failed commands, successful fixes, and rejected explanations.

The memory agent changes these records through predefined tool calls rather than freely rewriting the bank. It can then reactivate a stored state by producing a targeted reminder. Meta describes the module as compatible with existing agents and harnesses, rather than as a replacement for the primary model or a general-purpose strategic adviser.

What the reported benchmarks show

The system was evaluated on Terminal-Bench 2.0, which tests autonomous agents in realistic command-line tasks, and tau2-Bench, which measures conversational tool use in airline, retail, and telecom settings.

In the reported setup, Claude Sonnet 4.5 served as the action agent and Claude Opus 4.6 as the memory agent. The combined system completed 46% of Terminal-Bench tasks on its first attempt, compared with 38% for the baseline. On tau2-Bench, the task-weighted average rose from 55% to 62%.

The results were uneven. Airline and retail scores each increased by roughly 10 percentage points, while telecom improved by about 3 points. The researchers interpret that variation as evidence that the value of intervention depends on the task rather than following a universal reminder rate.

The reported gains remained when the stronger Opus 4.6 model was used as the action agent, although they were smaller: 2.4 percentage points on Terminal-Bench and 2.5 points on tau2-Bench. That result suggests the memory layer is not only compensating for a weaker primary model.

These are research results reported by Meta’s authors, not independent validation or evidence of production adoption. The paper also notes that the memory agent sometimes treated speculative inferences as more certain than they deserved. The remaining failures were therefore often about calibration, not simply whether relevant information had been saved.

Why selective recall matters for builders

For AI builders, the proposal addresses a practical weakness in agent workflows: task history is not the same as reliable task state. A longer context can preserve more text without ensuring that the model will use an earlier warning when a later decision conflicts with it.

The distinction matters in coding agents, support automation, and other systems that call tools repeatedly. A failed shell command should influence the next attempt. A verified customer record should outweigh an unsupported claim. A hard requirement should remain active even after the agent shifts attention to debugging or another subtask.

Meta’s ablation tests support that narrower interpretation. Giving the action agent the complete memory bank at every step reduced performance compared with selective reminders. Removing the silence option also produced less consistent results across domains. An adviser-style system without a persistent memory bank helped in some areas but hurt in others.

The design also outperformed Mem0 in the reported comparison. That difference, according to the research description, was not limited to retrieval quality. Meta’s memory agent decided whether a stored state should enter the agent loop and how it should be expressed as a reminder.

That could make the architecture useful where reliability matters more than maximum conversational recall. It also introduces another model call, another source of latency, and another place where incorrect judgments can affect the workflow. Teams would need to measure the cost of reminders against the cost of repeated actions, failed tool calls, and manual recovery.

Training and deployment questions remain open

The main version does not require a specially trained model; it uses prompting and tool-constrained updates. Meta also tested a smaller Qwen3.5-27B model as the memory agent while keeping a much larger action model fixed. Without additional training, the smaller model reduced performance. Supervised fine-tuning recovered the loss, and reinforcement learning later improved its decisions about when to recall a stored state.

That result complicates the idea that memory can be added as a simple plug-in. A prompted memory layer may work with capable models, but lower-cost deployments could need task-specific training to make intervention decisions reliably. The system’s fixed review schedule is another operational constraint: a future version might invoke memory when needed rather than checking at predetermined intervals.

Meta also identifies unresolved choices about whether literal records or more abstract task summaries work better, and whether the memory and action agents should be trained together. Those decisions could affect auditability, portability between models, and the ability to diagnose why an agent acted on—or ignored—a particular memory.

What to watch next

The most important follow-up is independent testing on additional agent tasks. The current evidence comes from two benchmarks and a research evaluation, so it is not yet clear how consistently the method transfers to software engineering, enterprise operations, or long-running browser workflows.

Builders should also watch for measurements of cost and latency, not only task success. A memory coach that improves completion rates but adds frequent model calls may be attractive for expensive failures and impractical for high-volume automation.

Further signals include adaptive invocation instead of fixed-interval review, stronger confidence calibration, and comparisons with other memory systems. Open implementations or reproducible evaluations would make it easier to determine whether the gains come from the two-agent structure, the selective reminder policy, the model pairing, or the task-specific prompting.

Creati.ai perspective

Meta’s proposal frames agent memory as a control problem, not merely a storage problem. The valuable capability is deciding when a previous state should change the next action, while avoiding a flood of reminders that makes the primary agent less effective.

That is a useful direction for teams building reliable AI agents, but the benchmark gains should be treated as an early research signal. The practical test will be whether selective memory reduces real operational failures after accounting for extra inference cost, mistaken reminders, and the difficulty of auditing a second agent’s intervention policy.