Two independent developer showcases describe new tooling aimed at reducing outdated context in AI coding assistants. Dev.to introduces “mnemoa,” which scans repository instruction files (such as AGENTS.md and CLAUDE.md) and project plans to find stale, redundant, broken, or contradictory guidance so it can be pruned or converted into tool-enforced checks. The focus is keeping repository memory “small, accurate, and useful” by removing instruction drift and signaling issues like invalid paths or commands.
A Hacker News Show HN post presents “Mnemolith,” an MCP server that provides shared, per-repository memory for multiple coding agents (Claude Code, Cursor, and Codex). It records which files memory depends on using git blob hashes; when those files change, memories are set aside for review rather than silently continuing. It also includes a local dashboard to view memory provenance and history, and supports syncing current memories back into AGENTS.md/CLAUDE.md for PR review.
Across both writeups, the different emphasis is on either pruning and promoting instruction content (mnemoa) or tracking dependencies so memories are flagged when underlying code changes (Mnemolith).
