AtomicMemory is a memory layer that gives agents durable context across sessions without tying an application to one model, framework, or deployment. Memory is correction-aware: as facts change, a memory is superseded, clarified, deleted, or retained rather than appended to an ever-growing recall log. Extraction, embeddings, mutation, reranking, and retrieval each sit behind an explicit provider boundary, so each piece is swappable.
AtomicMemory runs as a managed Hosted Cloud or as the open-source Core on your own machine (Connected Local), which can link to Cloud for trace visibility. Core is a Docker-deployable backend with mutation, retrieval, and Postgres/pgvector storage. Running Core yourself lets you audit the mutation and retrieval path instead of relying on a hosted black box. Connected Local uses an OpenAI API key for model calls.
One memory protocol is exposed through several surfaces: the am CLI, an MCP server that the CLI configures for agent hosts, a server-side TypeScript SDK, a Python SDK, framework adapters, and host plugins. A companion bridge imports llmwiki knowledge bases into AtomicMemory.
Reported benchmark results under matched methodology:
| Benchmark | AtomicMemory v66 | Cost per question |
|---|---|---|
| BEAM-100K lenient | 0.7375 | $1.26 |
| BEAM-1M lenient | 0.6625 | $0.083 |
| BEAM-10M lenient | 0.4875 | $0.081 |
| LoCoMo10 (GPT-4o-mini, binary) | 0.8396 | $0.066 |
Features
- Correction-aware memory: supersede, clarify, delete, or retain memories as facts change
- Hosted Cloud or Local: start on the managed service or run the open-source Core without changing the integration
- Inspectable Core: an auditable mutation and retrieval path backed by Postgres and pgvector
- Provider boundaries: swappable extraction, embeddings, LLMs, stores, and artifact storage
- CLI: the am command sets up profiles, ingests and searches memories, and diagnoses host configuration
- MCP server: memory tools for agent hosts such as Cursor, Claude Code, Codex, and OpenCode
- SDKs: a server-side TypeScript memory client and a Python SDK
- Framework adapters: AI SDK, OpenAI Agents SDK, LangChain, LangGraph, and Mastra
- Host plugins: Claude Code, OpenClaw, and Hermes, with Codex and Cursor plugins coming soon
- llmwiki bridge: imports an llmwiki JSON export while preserving its metadata
