Large language model agents can now autonomously work through repository-level coding tasks that require hundreds of think-act steps. Yet according to a new paper on arXiv, the experience each run produces is rarely carried over to the next run. This means agents effectively start from scratch every time, even when facing similar problems.
The paper introduces REM, a memory framework designed specifically for coding agents. It focuses on three core questions: what information should be stored, when it should be recalled, and what should be forgotten. By making these decisions explicit, REM aims to turn past runs into reusable knowledge rather than discarding them.
The abstract notes that prior memory approaches have been insufficient, though it does not detail them. The significance of REM lies in treating memory as a deliberate design component for long-horizon agent tasks, potentially improving efficiency and consistency in complex coding workflows.