AI agents forget between sessions. Bolt on a memory store and you inherit a quieter failure: when the store breaks, most of them hand back an empty result that looks exactly like "nothing found." Your agent can't tell "I remember nothing" from "I'm broken." aura-mnemos closes both gaps — it remembers, and when it can't, it says so.
pip install aura-mnemos"Survives failure" is easy to write and hard to prove. So the test suite causes the failure — deletes the database, corrupts the file, drops the table — and checks that health turns red and the tools return a clean error instead of crashing. A green test over a store that never broke would be a lie, so the tests break it first.
It's v0.1.0, shipped today — there's no adoption to brag about, and this page won't pretend otherwise. What's proven is that it works: installed from PyPI onto a clean machine, driven end-to-end, and put through 22 tests including the disaster cases above.
recall is substring search, not semantic search — simple on purpose.
A graph layer, native apps, and framework adapters are named as roadmap in the README,
not shipped. The obvious name mnemos was already taken on PyPI, so this is
aura-mnemos. No cleverness — just accuracy.
pip install aura-mnemos
# point your MCP client at the `aura-mnemos` command — then:
# remember("the deploy key rotates on the 1st") · recall("deploy key")
# and watch it stay honest when the store dies:
MNEMOS_DB=/nonexistent/db.sqlite aura-mnemos-health # /health → downFree and open. Install it, break it on purpose, and watch it tell the truth.