Five researchers propose a way to improve how software agents store what they learn. Agents that keep memory across sessions usually write notes from finished tasks alone. Those notes can lock in mistakes, overgeneralise, or go out of date. The new method gives the memory-writing agent read-only tools to check facts against the live environment. It needs no model retraining and leaves the task agent, retriever and write permissions untouched. Tests ran in a GitHub Copilot harness built on its software development kit. The authors used a database exploration benchmark and 90 management-consulting tasks. Pass rates, rewards and costs all improved, and tool calls fell across six task environments. Gains held on two Claude models without the memory format drifting. The work is a preprint and has not been peer reviewed.
What changed
Memory curation relied only on completed task records, which could store errors or stale facts.
What it unlocks
Letting a memory-curating agent check candidate memories against the live system before saving them.
- pass rate 39% to 73%
- reward 8.60 to 22.60
- cost $3.38 to $1.68 per task
- queries 8.8 to 4.7 per question
What you need to act on it
- an agent platform with persistent memory
- read-only access to environment tools
Sources