What is mememory?

mememory is a persistent semantic memory server for AI agents. It implements the Model Context Protocol (MCP) and gives agents the ability to store, search, and recall knowledge across sessions.

The Problem

AI agents start every session with a blank slate. They forget your preferences, project context, architectural decisions, and lessons learned. You end up repeating yourself.

The Solution

mememory stores knowledge as vector embeddings and delivers it back to the agent automatically at session start. The agent remembers your rules, your project context, and your preferences — without you asking.

How it works

Session starts
    ↓
mememory bootstrap loads your rules and context
    ↓
Agent has full memory from the first message
    ↓
During the session, agent stores new knowledge via MCP tools
    ↓
Next session — everything is remembered

Key Features

• Semantic search — recall by meaning, not keywords
• Hierarchical scopes — global rules and project-specific context, with project inheriting global
• Contradiction detection — warns when new memories conflict with existing ones
• Belief evolution — supersede old knowledge without losing history
• Auto-expiry — TTL for temporary context (sprint goals, deadlines)
• Session bootstrap — bootstrap-type memories loaded automatically at session start
• Pluggable embeddings — Ollama (local), OpenAI, or any compatible provider
• Privacy first — all data stays on your machine

Architecture

mememory CLI (host)           Docker stack
───────────────               ────────────
bootstrap ──────HTTP────────> Admin API (:4200) ──> PostgreSQL + pgvector
                                   │
                                Ollama (embeddings)

• mememory — native Go binary on the host. Handles setup, bootstrap, status.
• Docker stack — PostgreSQL with pgvector for storage, Ollama for local embeddings.
• MCP server — runs via stdio, connects agent to the memory store.