Meta AI pairs two agents to keep long tasks on track with targeted memory

The core problem Meta identified is that as task history grows, the behavioral state guiding an AI agent's decisions gets scattered and may become buried in the context window or fall out entirely. Even when the information technically remains available, it no longer reliably shapes the agent's behavior. This is distinct from personalization or recall challenges that existing memory systems handle well; agents working through a multi-step task need the system to actively decide *when* a past decision or constraint should re-enter the current work. Existing approaches—either constant recall (which adds latency and token overhead) or fixed summarization rules—fail because failure modes vary widely by task, so no single aggregation rule can reliably decide when intervention helps versus distracts.

Meta's design addresses this by separating concerns: a "memory agent" (prompted with no special training required) maintains a structured bank that captures what the action agent has tried, what happened, and what constraints apply, while an "action agent" remains unmodified and continues its work. The innovation is not storage but *selective intervention*—the memory agent has a "silence" option and only reminds when it believes memory should influence the next step. Testing showed this selective approach outperformed both always providing full memory and a production memory system (Mem0) that retrieves records through search. The gains persisted across model scales: with weaker Claude Sonnet 4.5 as the action agent, task completion rose from 38 to 46 percent on command-line benchmarks and from 55 to 62 percent on conversational benchmarks; with the stronger Opus 4.6, improvements of 2.4–2.5 percentage points still appeared, suggesting the memory system compensates for capacity constraints rather than merely masking a weak baseline.