atomicstrata/atomicmemory

AtomicMemory is a memory layer for AI agents. When you build an app that talks to a chatbot like Claude or GPT, the chatbot has no built-in long-term memory of past conversations or facts about the user. AtomicMemory gives the app a place to save that context, search it later by meaning rather than by exact words, and feed the right bits back into the next AI call. The README describes it as portable and inspectable, in contrast to hosted memory services that act as a black box around the data. The project ships several pieces. The Core is a backend that stores memories in Postgres with the pgvector extension for semantic search, and runs from a Docker image. The TypeScript SDK is the main programming interface, with helper functions for capturing messages, searching by query, and packaging context. There are also a command-line tool, an MCP server (a protocol used by Claude and similar tools to call external services), framework adapters for Vercel AI SDK, OpenAI Agents SDK, LangChain, LangGraph, and Mastra, and plugins for agent hosts such as Claude Code, OpenClaw, Hermes, Codex, and Cursor. The README emphasises that memory is not just appending and recalling. The system supports supersession (newer facts replacing older ones), clarification, deletion, no-op decisions when nothing changes, and a record of how each fact was reached. The Core separates the pieces that handle embeddings, extraction, mutation, reranking, retrieval packaging, and evaluation, so the engine is not tied to one specific model version. It can run locally on a developer laptop for privacy-sensitive work, or against a hosted service. The README publishes benchmark numbers against industry tests called BEAM and LoCoMo10, claiming parity or better with named competitors at lower cost per query. The authors are careful to say these are scores under matched methodology rather than marketing claims, and that latency and recall claims need to be paired with the hardware, dataset, and date used to produce them. Performance is described as a design target of single-digit-millisecond local retrieval, not a guarantee. Getting started is shown as installing the npm packages, pointing a MemoryClient at a local Core instance, calling ingest to add messages, and calling search to retrieve them by query and user scope. The repository excludes the hosted service infrastructure, the Python SDK, and the raw benchmark research code. The license is Apache 2.0.