A new way to help agents remember work for future tasks
The CoReMem framework helps agents decide what information is worth saving to reduce effort in upcoming tasks.
Researchers created a framework called CoReMem to help agents handle a sequence of tasks. These tasks happen on a repository that changes over time. Each new task starts with the state left by the previous one.
Before, memory methods mostly focused on the current task. Other methods shared general experience across different repositories. Some methods saved memory but reset the repository state between tasks. This made it hard to know what information would be useful for a future task that had not arrived yet.
CoReMem uses prospective segment valuation to solve this. It looks at how much a piece of information saves effort on the next task. This helps create a compact memory using anchors. Anchors are specific locations in a repository or executable checks.
The framework also uses on-policy adaptation to improve how an agent uses memory. It uses provenance-gated value credit to reward specific steps. This credit goes to steps that use anchors provided by the memory. This helps the agent learn to use memory only when it actually saves effort.
In tests on the ChainSWE benchmark, CoReMem improved results. A 27B actor model saw its resolve rate increase from 21.0% to 25.7%. The model also used 17% fewer steps to finish tasks.
The system was more efficient with data. It used 11% fewer total cache-aware tokens compared to an untrained model without memory. On average, CoReMem raised the resolve per token by 41% across three different actor models.
The framework aims to balance the benefit of remembering against the cost of reading and writing that memory. It seeks to lower the effort of unseen tasks without losing the ability to solve them.
Key facts
- CoReMem increased the resolve rate of a 27B actor from 21.0% to 25.7% on ChainSWE.
- The framework achieved 17% fewer steps and 11% fewer total cache-aware tokens.
- CoReMem uses anchors like repository locations and executable checks to build compact memory.