kadubon/certified-memory-governance-layer

CMGL, short for Certified Memory Governance Layer, is a Python library designed to act as a gatekeeper for AI agents that need to store and retrieve memories over long-running sessions. Think of it as a strict checkpoint that sits between an AI agent and whatever database it uses to remember things. Before any piece of information is written to memory, CMGL checks whether it is allowed, records evidence of that decision, and issues a typed receipt explaining why it was admitted or blocked. The core idea is procedural admissibility. CMGL does not judge whether a remembered fact is true, but it does enforce rules about the process of writing it. It can block stale, contradicted, unauthorized, or contaminated memory items, for example keeping AI-generated summaries out of the factual memory store, or requiring that writes carry a structured authority bundle proving they were permitted. Every decision is logged in an append-only ledger, a file where records are added but never changed, and each entry gets a cryptographic digest so the ledger can be audited later. CMGL runs entirely offline. It requires no API keys, no cloud services, and makes no LLM calls for its core checks. It connects to several popular agent memory frameworks, including Mem0, Graphiti, LangMem, and LangGraph, through adapter shims that the library provides, so you can add governance to an existing system without rewriting it. Installation is available through PyPI with a package manager called uv, and a command-line tool lets you initialize a local ledger, create authority bundles, and write governed memory entries in about ten minutes. The tool is aimed at developers building production AI agents where auditability and memory integrity matter. The full README is longer than what was shown.