An open-source project called projectmem wants AI coding agents to stop forgetting what already failed.
The tool, described in a new arXiv paper, keeps a local, append-only log of typed events - issues, attempts, fixes, decisions, and notes - from a coding project. Before an agent edits a file, it can query that log through the Model Context Protocol and get a precheck of recorded failures, open issues, and churn tied to that specific file. The author ran a six-month self-study across two machines and 27 projects, logging 3,228 events. In 86 of 427 closed issues, the eventual fix followed at least one earlier failed attempt. On a synthetic 1,500-event repository, a project analysis that took 42 seconds dropped to 79.3 milliseconds, with git subprocess calls falling from 1,500 to two.
The real bottleneck for coding agents usually isn't the model - it's that each session starts from zero, with no record of what the team already tried and ruled out. Persistent, file-scoped memory is a plausible fix for a problem every agent user has hit: the same busted approach suggested twice. The speed numbers also matter practically, since nobody wants an assistant that pauses for 42 seconds to remember what it did yesterday.
Worth noting: the paper itself is careful to say its failure histories are candidates for future warnings, not proof that warnings actually stop repeat mistakes. That distinction is the honest part most launch posts skip, and the authors say a controlled benchmark is still needed to show the memory layer changes outcomes rather than just logging them.