AtomicMemory

AtomicMemory gives agents and AI applications durable memory across sessions without tying the application to one model, framework, or deployment. Memories are correction-aware: facts can be superseded, clarified, deleted, or retained as they change, rather than accumulating as append-only recall. The same memory protocol is exposed through a CLI, an MCP server, a TypeScript SDK, framework adapters, and host plugins.

There are three entry paths. Hosted Cloud is a managed runtime set up by am init, needing no Docker or OpenAI key. Connected Local runs the open-source Core on your machine — Docker, an OpenAI API key, and macOS or glibc Linux on x86_64 or arm64 — and can link to Cloud for trace visibility. The server-side TypeScript SDK talks to Core directly through a MemoryClient with ingest and search calls scoped per user. Core stores data in Postgres with pgvector and keeps extraction, embeddings, mutation, reranking, and retrieval packaging behind explicit provider boundaries.

Agent hosts are configured with am integrate --host, which writes the host's user-level MCP configuration; marketplace plugin packages are published for Claude Code, OpenClaw, and Hermes, while Codex and Cursor packages are listed as coming soon. Framework adapters are published for Vercel AI SDK, OpenAI Agents, LangChain, LangGraph, and Mastra. The CLI installer verifies every download against SHA256SUMS and can require GitHub build-provenance attestation.

The project reports the following results for version 66 under matched methodology:

Benchmark	Score	Position	Cost per question	Sample
BEAM-100K lenient	0.7375	Parity with Hindsight at 0.75	$1.26	n=80
BEAM-1M lenient	0.6625	+0.022 versus the Mem0 paper	$0.083	n=80
BEAM-10M lenient	0.4875	Parity with Mem0-new at 0.486	$0.081	n=80
LoCoMo10 GPT-4o-mini binary	0.8396	+0.171 versus the Mem0 paper	$0.066	n=1540

Features

• Correction-aware memory: supersede, clarify, delete, or retain memories instead of only appending
• One protocol, many surfaces: CLI, MCP server, TypeScript SDK, framework adapters, and host plugins share the same contracts
• Cloud or local: managed Hosted Cloud or the open-source Core on your own machine, with the same integration surface
• Core storage: Docker-deployable backend with mutation, retrieval, and Postgres with pgvector storage
• Provider boundaries: extraction, embeddings, mutation, reranking, and retrieval packaging are explicit and swappable
• Host integration: am integrate writes user-level MCP configuration; plugins are published for Claude Code, OpenClaw, and Hermes
• Framework adapters: published packages for Vercel AI SDK, OpenAI Agents, LangChain, LangGraph, and Mastra
• Scoped SDK calls: ingest and search take an explicit user scope, with a separate client for the storage namespace
• Credential handling: project-bound, per-installation keys stored with owner-only permissions and never printed
• Installer verification: SHA256 checksums always, with optional required GitHub build-provenance attestation
• Python SDK: a separate atomicmemory package is published on PyPI