Give all your AI coding tools one shared memory (free, local, 3 commands)

A developer built Aletheia, a free, local memory and rules hub that gives multiple AI coding tools a single shared store of project decisions and rules instead of separate CLAUDE.md, AGENTS.md and Cursor rules files. It runs on Bun 1.1+, listens only on 127.0.0.1 with no account or telemetry, redacts secrets before storage, and wraps retrieved memory in a marked block so models treat it as data rather than instructions. The project is in alpha: the memory engine and installers pass their tests, but the developer has not yet opened a real host app with a generated config and measures retrieval rather than answer accuracy.

If you use more than one AI coding tool, you have probably noticed two things. First, every tool forgets. Close the session and the decisions, the quirks of your codebase and the "never do it this way" rules are gone. Second, every tool wants its own rules file: CLAUDE.md , AGENTS.md , Cursor rules, and so on. You end up copying the same rules into several places, and they drift apart. I wanted one place for both. So I made Aletheia : a free, local memory and rules hub for AI coding tools. No account. No telemetry. It listens on 127.0.0.1 only. You need Bun https://bun.sh 1.1 or newer. npm install -g aletheia-mem aletheia-server open http://127.0.0.1:37790 That is the whole setup. There are no runtime dependencies to install. Aletheia can write the config for the tools it knows. By default it only shows you what it would do: aletheia install --list what it can set up, per tool aletheia install claude-code dry run: prints every file it would change aletheia install claude-code --write Without --write nothing is written. With it, every changed file gets a timestamped backup first. Secrets stay out. Before anything is stored, capture hooks redact secrets and anything inside <private blocks. Memory is treated as untrusted. A stored note could contain text like "ignore your instructions". So when Aletheia hands memory back to a model, it wraps it in a marked block and tells the model it is data, not instructions. Tags inside notes are neutralised so a note cannot close the block early. It is alpha . The memory engine and the installers pass their tests. But I have not yet opened a real host app with a generated config, so I do not claim it "works with" any specific tool. I also measure retrieval, not answer accuracy, and I say so in the README. I built it with Claude Code: agents did most of the implementation in separate git worktrees, one task per commit, and I wrote the specs, reviewed the results and decided what counted as done. The most useful thing you can do is try it with the tool you use and tell me what happens. Open an issue with the tool, its version, the command you ran and what the tool did. What would you want your tools to remember?