rcarmo/memento

Memento is a shared memory system built so that several separate AI agents can know the same facts without sharing their private chats, reminders, passwords or machine specific notes. The author built it because they run multiple agent instances, one for personal work, one for servers, and others for individual projects, and all of them sometimes need to know the same background facts, like where a service runs or why one system replaced another. Memento works by giving these agents an authenticated connection, using the MCP protocol, to a shared collection of Markdown documents called concepts. An agent can search this collection, read a concept, follow links between related concepts, and propose a change. A human curator then reviews proposed changes before they are saved permanently to a Git repository. Every concept has a stable identifier, structured metadata, and normal Markdown links, so it can be read directly in a text editor or tracked through Git history. What belongs in Memento are facts meant to outlive a single conversation and be useful to more than one agent, such as who owns a service, how projects and machines relate to each other, or reviewed operating knowledge several agents should follow. Personal chat logs, daily notes, reminders, and any passwords or tokens are meant to stay with the individual agent or machine instead. Searching uses an exact text search method by default, with an optional smaller AI model that can add meaning based ranking so different wording for the same idea can still be found. A separate small fine-tuned model can help route simple natural language requests to the right search action, which Memento then checks before running. There is also an optional visual debugging view that shows how memories and their links are organized, meant only for a trusted development setup rather than a public facing service. Memento runs as a container and supports recent versions of Python. It is aimed at developers already running multiple AI agents who want those agents to share durable, reviewed knowledge instead of duplicating it or losing it between sessions. The project is MIT licensed, so it can be used freely including for commercial purposes.