inspeximuswe have inspected

mcp server

Your agent forgets what it decided. One line fixes that.

An MCP memory server for Claude Code, Cursor, Windsurf, Codex and Cline. The next session starts knowing what the last one decided — and says so before the action that would contradict it, not in the post-mortem.

pip install inspeximus
inspeximus install --ide claude     # also: cursor, windsurf, codex, cline

No database to run, no API key to supply. The core is one zero-dependency file; the MCP server is an opt-in extra.


what gets wired

Three hooks, and the moment each one fires

Memory that only answers when asked is memory you have to remember to consult. These fire on their own.

  1. SessionStart — injects the decisions still in force. Not everything ever said: what has not been superseded. A later correction retires the earlier decision by topic, so the new session inherits a current position rather than an archive.
  2. PostToolUse — captures what actually happened, keyed by file. The record is what the run did, not what the plan said it would do.
  3. PreToolUse — surfaces the decision that bears on the action before it runs. This is the one that changes outcomes: a rule you are reminded of after the commit is a rule you already broke.
Built because we needed it. The PreToolUse guard exists because a rule we had agreed on was given twice and broken twice in the same day — by us. A rule you have to remember is a rule you will break at the moment you feel most certain.

the surface

68 tools any MCP host can call

Recall and write are the two you will use daily. The rest are there because agent memory that is only additive becomes wrong and stays wrong: correction, revert, supersession history, provenance checks, contradiction detection, subject erasure with receipts, and audit bundles a third party can verify without trusting the store.

The count is read from the server’s own registrations by our claims audit, not typed into this page — a published count maintained by hand drifts on exactly the schedule you stop watching it, and this one did, twice, before it was automated.


questions

Reasonable objections

How is this different from notes in CLAUDE.md?

CLAUDE.md is a file you maintain by hand, and it grows until it gets skimmed. This captures decisions as they happen and keys them by topic, so a later correction retires the earlier decision instead of sitting beside it. You never prune it, and you never have to wonder which of two contradictory lines is current.

Will it slow my session down?

Recall is local and lexical by default — no network call, no embedding service, no model in the loop. Semantic recall is available if you supply an embedder, and optional because most sessions do not need it.

What happens when a stored decision turns out to be wrong?

You correct it, and the old one stops coming back. That is the property the whole library is built around, and we measured it against other memory systems rather than asserting it.

Is my data going anywhere?

No. The store is a file on your machine. Encryption is an opt-in extra, and erasure is a real operation with a receipt rather than a filtered view.