CoReMem trains coding agents to use memory on the next repository task
CoReMem evaluates memory by what it contributes to a coding agent’s next task on an evolving repository. A preprint combines concrete file and command references with training for the model that reads them. Its own experiments report higher task resolution and lower resource use while including the cost of writing and reading memory.
Artificial Intelligence··Night
Memory is scored on an unseen next task
CoReMem, a proposed memory framework for coding agents, evaluates stored information against a future task in the same repository. Earlier changes remain in place, but the next assignment is unknown when memory is written. The preprint compares runs starting from the same inherited repository state. In the authors’ ChainSWE experiments, a model with 27 billion parameters achieves a 25.7 percent resolution rate, up from 21.0 percent. The comparison includes the resources used to write memory.[1]
File locations and commands guide later work
The researchers divide the earlier task trajectory into seven functional segments and remove them in turn to measure their contribution. Commands and repository content provide reusable information; extended reasoning narratives deliver less of the measured savings. The resulting memory emphasizes locations in files, identifiers in code, commands for testing and their exit codes. The experiments also examine replacing full diffs with a list of changed files, reducing the representation of earlier work.[1]
Training rewards steps that use memory
CoReMem also trains the model that reads the memory, comparing its own memory, no memory and memory from another repository. Successful resolution takes priority; step costs then distinguish successful runs. Additional credit goes to steps linked to a memory-supplied location or executable check. Other steps retain their ordinary outcome reward. With separate training and test repositories, the authors report a 17 percent reduction in steps and an 11 percent reduction in cache-aware token totals, compared with an actor lacking both training and memory.[1]