An agent’s memory should include the right moment to act

Consider an assistant asked to prepare a briefing when a colleague sends a revised document. Remembering the request is only one part of the job. The assistant must notice the revision, recognize that it satisfies the condition, and check whether the user has since cancelled the briefing. A system can recall the original conversation perfectly and still disappoint at every later step.

PM-Bench, introduced by UCLA researchers Genglin Liu and Saadia Gabriel on July 14, 2026, makes this problem explicit. It evaluates prospective memory: maintaining an intention and acting when a future condition becomes true. Its simulated week combines ongoing activities with delayed obligations, rescheduling, cancellations, and information that requires active checking. The experiments compare eight models across eight configurations. Their broader finding is a trade-off: more aggressive monitoring can recover missed tasks while also producing unnecessary actions. No configuration consistently solves every dimension. PM-Bench, v1 ↗

The accompanying repository makes the experiment inspectable. It includes the deterministic v9 scenario, scoring machinery, configurations, and reported trajectories. Two configurations are replay-based voting ablations, a distinction worth preserving when comparing them with live model runs. PM-Bench repository ↗

This work connects to an architectural distinction in OpenAI’s May 1, 2026 cookbook on memory and compaction. In its synthetic evidence-review example, compaction carries the current working state through a long interaction, while persistent memory carries reusable workflow lessons into later runs. The reviewed memo remains the authoritative investigation artifact. The example deliberately limits memory to habits and preferences rather than storing case conclusions as reusable truths. That separation gives developers a concrete way to avoid letting a useful summary quietly become an unexamined source of authority. OpenAI cookbook ↗

Anthropic’s September 29, 2025 context-engineering guidance provides another piece. It describes keeping durable notes outside the active context and retrieving relevant information when needed. It also warns that aggressive compaction can discard details whose importance becomes apparent later. The engineering objective is selective continuity: preserve consequential state while reducing irrelevant material. These are design techniques, with trade-offs, rather than guarantees that an agent will remember correctly. Anthropic engineering ↗

A practical interpretation is to represent a future commitment as a small, inspectable record. It should identify the requested outcome, the trigger, the source of the request, and its current status. A rescheduled deadline should supersede the earlier one. Completion should prevent duplicate execution. A missing update should remain unknown instead of being silently treated as confirmation. These are proposed design requirements, not findings that any particular product has already satisfied.

Evaluation should then exercise the full lifecycle. Give the agent an intention, interrupt it with unrelated work, change one condition, and observe the next action. Include cases where the correct behavior is waiting. Track missed obligations, premature actions, repeated actions, and unnecessary checks separately. A single aggregate score can hide a system that achieves good recall by interrupting the user too often.

The limits matter. PM-Bench uses synthetic scenarios and is not a readiness certification for high-stakes deployment. Its authors explicitly caution against translating benchmark performance into unsupervised use. The benchmark is most useful as a diagnostic lens for a failure mode that ordinary question-answering tests can miss. PM-Bench ethics statement ↗

For PersonaSI research, the useful next question is how to make commitments understandable and correctable by the person who made them. A remembered instruction should carry its history, current conditions, and boundaries. This article identifies a research direction; it does not describe features shipped by PersonaSI.