One Memory Layer, Multiple Models (Claude, GPT, Llama, etc.) (github.com)

MemMachine is an open‑source, model‑agnostic memory layer for AI agents that persistently records and recalls user data across sessions, agents, and large language models (GPT, Claude, Llama, etc.). It exposes developer-friendly APIs (Python SDK, RESTful, MCP) and ships as a Docker container and Python package under Apache 2.0, making it easy to integrate into assistants, autonomous workflows, and research agents. The system distinguishes multiple memory types—Working (short term), Persistent (long term), and Personalized (profile)—and organizes stored information as Episodic Memory (conversational context) in a graph database and Profile Memory (long‑term facts) in an SQL store. This matters because persistent, structured memory bridges the gap between one-off chatbots and genuinely personalized, context-aware assistants: agents can remember preferences, histories, and relations over time to reduce repetition, improve relevance, and enable task continuity. For researchers and engineers, MemMachine’s architecture supports experiments in agent cognition and multi-agent workflows while offering real-world use cases (CRM, healthcare continuity, finance advice, consistent content writing). Technical implications include richer retrieval strategies via graph‑based episodic links, scalable profile querying in SQL, and cross‑model interoperability that lets different LLMs access a unified user profile. The project is community-driven (GitHub issues, Discord) and includes docs, a Quick Start, and examples to accelerate adoption.