[ about ]

Memory for coding agents should be checked, not trusted.

Coding agents now do work that spans many sessions: one session plans, another implements, a third is started by a script at night and picks up where the last one stopped. Whatever they remember between those sessions is only useful while it is still true of the code. memprove exists to answer that one question, with evidence instead of a guess.

It is open source under the Apache License 2.0 and runs on your own machine. Version 0.1.0 was released on 2026-11-14.

How it is built

• Evidence over wording. A fact is verified only while a test run it cites passed in the repository state you have now. memprove never judges the sentence itself.
• Deterministic. Verdicts come from fixed rules, not a model. The same ledger, repository state and budget give the same bytes.
• Local. The daemon talks over a Unix socket and has no network access. There is no account and nothing to send home.
• Nothing is deleted. A statement that no longer holds is superseded or withdrawn, and the record says which Session did it.
• Honest limits. No tests, no verified. The docs list what memprove cannot see, such as ignored files and environment variables.
• Set up by your agent. llms.txt gives a coding agent the steps, so integration happens in the session you already work in.

Where this is going

memprove is the first tool of a larger effort: software that makes work done with AI agents verifiable, so that a person can trust a hand-off without re-reading everything. Each tool that follows will be held to the same rule as this one: show the evidence, say what you cannot see.

Get in touch

Questions, bugs and ideas go to GitHub issues. Releases and their notes are on GitHub Releases and in the changelog.