Why AtomicMemory
Most memory products ask you to trust a hosted black box with the layer that decides what an AI believes about your users. AtomicMemory takes the opposite position: the interface should be portable, the engine should be inspectable, and the memory system should be able to revise itself when facts change. Memories are not just appended and recalled — the engine makes explicit mutation decisions (add, update, supersede, clarify, delete, or no-op) so your memory store stays accurate over time, even as users change their minds.Benchmark results
AtomicMemory v66 reaches leading performance/cost results under matched methodology on three independent benchmarks, and matches the strongest published result on a fourth.| Benchmark | Score | Position | Cost / Query |
|---|---|---|---|
| BEAM-100K lenient | 0.7375 | Parity with Hindsight at 0.75 | $1.26 |
| BEAM-1M lenient | 0.6625 | Leading performance/cost; +0.022 vs Mem0 paper | $0.083 |
| BEAM-10M lenient | 0.4875 | Parity with Mem0-new at 0.486 | $0.081 |
| LoCoMo10 GPT-4o-mini binary | 0.8396 | Leading performance/cost; +0.171 vs Mem0 paper | $0.066 |
What’s included
AtomicMemory ships as a coordinated set of surfaces. Every surface talks to the same memory store through the same provider interface — swapping from a direct SDK call to a LangGraph adapter to a Claude Code plugin does not require re-ingesting your data or changing your retrieval semantics.Core
The open-source, Docker-deployable memory backend. Durable context, semantic retrieval, AUDN-SC mutation, Postgres + pgvector storage, and pluggable embeddings and LLM providers.
SDK
Backend-agnostic TypeScript client. Pluggable providers (
atomicmemory, mem0), local embeddings, storage adapters, and semantic search primitives. The portability contract every adapter is built on.CLI and MCP server
Command-line and MCP surfaces for setup, diagnostics, memory capture, retrieval, and context packaging. Use the CLI interactively or in scripts; the MCP server connects any MCP-compatible agent host.
Framework adapters
First-class integrations for LangChain, LangGraph, Vercel AI SDK, OpenAI Agents SDK, and Mastra — all built on the same SDK, not on separate implementations.
Host plugins
AtomicMemory also ships plugin surfaces for agent hosts. Published plugins are available today for Claude Code, OpenClaw, and Hermes. Plugins for Codex and Cursor are implemented and coming soon.MCP tools
The MCP server exposes four tools that any MCP-compatible host can call:memory_search, memory_ingest, memory_package, and memory_list.
Local-first, no lock-in
Core runs locally on a developer laptop (default port17350) or in your own infrastructure. The hosted profile is available where it makes sense. There is no capability cliff between the two — switching is a two-value config change (apiUrl and apiKey). Package APIs are stable and semver-disciplined, and migrating between SDK use, adapters, and host plugins does not require re-ingesting your data.
AtomicMemory Core is Apache 2.0 licensed and open source. The SDK, CLI, MCP server, framework adapters, and host plugins are all published on npm under the
@atomicmemory scope.Get started
Quickstart
Run Core in Docker, install the SDK, and ingest your first memory in under five minutes.
SDK Overview
Explore the full TypeScript client surface: providers, memory operations, storage, and embeddings.
API Reference
Browse the HTTP API endpoints for ingest, search, retrieval, and memory management.
Core Concepts
Understand scope, ingestion modes, AUDN-SC mutation, retrieval, and providers before you build.