Mwe-MCP

Most agent memory lives inside a single session, inside a single model, invisible to anything else running in your stack. mwe-mcp is a self-hosted Model Context Protocol server that exposes a shared Markdown wiki as persistent memory — agents read from it and write to it through standard MCP calls, and the files sit on disk where any human can open them. The core workflow is: agent makes an MCP call to read or store a fact, the server writes or retrieves a structured Markdown fragment, and every other MCP-compatible agent on the same server sees the same state.

The differentiating feature is the human-readable, file-based store combined with per-fragment access control lists. Unlike vector stores or opaque database backends, every memory entry is a Markdown file you can audit, edit, or delete directly. The ACL system lets you assign read and write permissions at the individual fragment level, so multi-user and multi-agent deployments can share infrastructure without sharing every secret — a design that matters when you’re mixing personal and professional agents on the same host.

The tool fits teams or individual developers who want transparent, governable agent memory and are willing to run their own server. It does not fit teams that need a hosted endpoint, an SLA, or a managed upgrade path. The AGPL license also means any product you ship that links against the server inherits that license — legal review is required before embedding it in a commercial product. The overnight self-organizing pass that restructures the wiki is described in the docs, but the behavior under large or rapidly growing wikis is not detailed in the available source material.