Analysis updated 2026-05-18
Give an AI agent a memory store that can correct a fact and keep the correction from being overwritten later.
Revert a stored fact back to its previous value with a single command.
Erase everything related to a subject from an agent's memory and get a receipt of the deletion.
Add an MCP-compatible memory server to an agent project without a language model on the write path.
| dancenitra/inspeximus | 1ncendium/aibuster | aaronmayeux/ha-hurricane-tracker | |
|---|---|---|---|
| Stars | 5 | 5 | 5 |
| Language | Python | Python | Python |
| Setup difficulty | easy | moderate | easy |
| Complexity | 3/5 | 3/5 | 2/5 |
| Audience | developer | ops devops | general |
Figures from each repo's GitHub metadata at analysis time.
inspeximus is a Python library and MCP server that gives AI agents a memory that can be corrected and erased on purpose, rather than one that just keeps accumulating facts forever. You install it with pip, or as a plugin inside Claude Code with a single command, and it stores what it remembers in a local file inside your project. The core idea the author points to is that most agent memory tools call a language model every time something is stored, to extract or summarize facts. inspeximus does not do this. Storing a fact is a plain, deterministic operation with no model call involved, which the README says makes storage free and repeatable, at the cost of not doing any of that automatic extraction work itself. Because storage does not depend on a model, the library can offer a few specific guarantees. Writing a new value for a fact replaces the old one, and a feature called echo_guard is meant to stop an old, corrected value from resurfacing later. A revert command can roll a corrected fact back to what it was before. A forget function is meant to remove a value entirely, including related records that were derived from it, and produces a receipt of the deletion rather than just silently wiping it. The project includes its own audit scripts that download the published package from PyPI and test these claims against the real released version, including deliberately trying to break the checks to confirm they can actually fail. The README documents at least two real bugs the audit process caught and fixed in past versions. The author is explicit that no formal certification exists for this kind of software, and that the receipts prove a deletion happened, not what the deleted content was. The project is released under the MIT license. The full README is longer than what was shown.
A Python library and MCP server that gives AI agents a memory they can correct or erase on command, without using a language model to store facts.
Mainly Python. The stack also includes Python, MCP, PyPI.
Use freely for any purpose, including commercial use, as long as you keep the copyright notice.
Setup difficulty is rated easy, with roughly 5min to a first successful run.
Mainly developer.
This repo across BitVibe Labs
Verify against the repo before relying on details.