A fragment-based long-term memory server built on the Model Context Protocol, supporting seven atomic memory types, semantic search, spreading activation, and reconsolidation for AI Agent cross-session knowledge persistence.
Memento MCP is a long-term memory server for AI Agents built on the Model Context Protocol (MCP), currently at v3.1.1. Its core design decomposes information into seven atomic "fragment" types for persistent storage and cross-session retrieval: fact, decision, error, preference, procedure, relation, and episode.
The system employs a 3-layer search architecture (Redis cache → PostgreSQL keyword → pgvector semantic search) and implements cognitive-science-inspired memory mechanisms including Spreading Activation, Reconsolidation, and Episode Continuity. Automatic maintenance covers deduplication, contradiction detection, importance decay, and TTL-based forgetting. Additional features include CBR (Case-Based Reasoning) mode, a Symbolic Verification Layer, affective tagging (6 labels), 4 operation mode presets, and an admin console with knowledge graph visualization.
Runtime dependencies include Node.js 20+, PostgreSQL 14+ (pgvector), and optional Redis for caching. Embedding supports OpenAI API or local Xenova/multilingual-e5-small model. It connects to major MCP-compatible platforms (Claude Code, Cursor, Windsurf, GitHub Copilot, ChatGPT) via Streamable HTTP or OAuth (RFC 7591). Achieves 88.3% search recall@5 and 45.4% QA accuracy on the LongMemEval-S 500-question benchmark. Licensed under Apache-2.0.