Running in under a minute
No API key, no account, no service, no LLM. Python 3.9 or newer and nothing else — inspeximus has zero required dependencies.
1. Install
pip install inspeximus
2. Correct a fact, and watch the correction hold
Every memory library can store and retrieve. This is about the fact that turns out to be wrong, or that was true on Monday and outdated by Friday.
from inspeximus import Inspeximus
m = Inspeximus("memory.json")
m.remember("The staging database is db-3.internal", key="staging-db")
m.remember("The staging database is db-7.internal", key="staging-db") # a correction
m.recall("which staging database")[0]["text"]
# 'The staging database is db-7.internal' <- the correction wins
m.revert("staging-db") # and it is reversible
m.recall("which staging database")[0]["text"]
# 'The staging database is db-3.internal'
Measured from a clean directory, three runs: 187, 190 and 193 ms end to end, including the import. No network, no model, no service.
3. Say the old value again. It still does not come back.
This is the part a most-recent-wins rule cannot do.
m.remember("The staging database is db-3.internal", key="staging-db")
m.recall("which staging database")[0]["text"]
# 'The staging database is db-7.internal' <- still the correction
m.remember("The staging database is db-3.internal", key="staging-db", reaffirm=True)
# going back is a decision you make on purpose
4. Or skip Python entirely
One line wires an MCP server and three hooks into your editor.
inspeximus install --ide claude # also: cursor, windsurf, codex, cline
From the next session on, your agent starts knowing what the last one decided. What each hook does.
5. Check us without trusting us
Neither command needs a key, a service, or any data of ours.
python claims_audit.py
Forty seconds. It reads every number we publish and reports whether each is registered, whether its pin still resolves, and whether a committed command recomputes it — including the rows marked WITHDRAWN, which are figures we published and then could not reproduce, kept beside the probe that refutes them rather than deleted.
python probes/integrity_bench_revert.py --systems inspeximus --judge local --n 5
Free, offline, deterministic, and it prints its own caveat that a local judge is not comparable with the OpenAI-judged figures on the measurement page. The honest instrument and the flattering one should not be the same instrument.