atomicstrata/atomicmemory
Free面向 AI 代理的可移植语义记忆，包含核心引擎、TypeScript SDK、框架适配器、MCP 服务器、CLI 和宿主插件，让记忆在不同工具间复用。
FreeFree tier
About atomicstrata/atomicmemory
AtomicMemory is a portable, inspectable semantic memory layer for AI agents and applications. It provides a core engine, TypeScript SDK, framework adapters, MCP server, CLI, and host plugins, allowing developers to embed memory directly where their AI code runs. AtomicMemory enables capturing context, grounding generations in prior interactions, and carrying knowledge across sessions. It is local-first where supported, hosted where convenient, and designed so the choice can change later without rewriting the application. The engine is open source and self-hostable, with benchmark results showing leading performance/cost on BEAM-100K, BEAM-1M, and LoCoMo10.
Key Features
Portable memory protocol consumed by direct SDK calls, CLIs, MCP server, framework adapters, and host plugins
SDK-agnostic: adapters are built on the same SDK, allowing drop-down to the SDK without data loss
Inspectable and open source core, enabling self-hosting and transparency
Local-first where supported, hosted where convenient, with the ability to switch without rewriting
Supports multiple frameworks including LangGraph, Claude Code, and Vercel AI via adapters
Benchmark-leading performance/cost on BEAM-100K, BEAM-1M, and LoCoMo10
Pros & Cons
Pros
- Open source and self-hostable, giving full control over data and operations
- Portable across SDKs, CLI, MCP server, and multiple framework adapters
- Inspectable memory engine allows debugging and fact revision
- Benchmark results show strong performance/cost ratio compared to competitors
- Local-first design reduces latency and dependency on external services
Cons
- Primarily focused on JavaScript/TypeScript ecosystem, limiting language support
- Documentation (docs.atomicstrata.ai) not directly visible in public repo; may require additional resources
- Benchmark coverage is limited to specific datasets (BEAM, LoCoMo); absence on BEAM-10M frontier noted
- Relatively new project (v66) with evolving roadmap and community
- No dedicated hosted service tier documented; self-hosting required for production use
Best For
Capturing context from AI agent interactions across sessionsGrounding generations in prior user interactions for consistencyCarrying knowledge between different AI tools and frameworks (e.g., LangGraph, Claude Code, Vercel AI)Building portable memory for multi-agent systemsEnabling inspectable memory in applications that require transparency and reproducibility
FAQ
Is AtomicMemory free to use?
Yes, AtomicMemory is open source and free to use. The repository is publicly available on GitHub under an open source license.
Does AtomicMemory work with LangGraph agents?
Yes, AtomicMemory provides a framework adapter for LangGraph, allowing the same memory store to be used across LangGraph agents and other tools.
What benchmarks have been published for AtomicMemory?
AtomicMemory v66 has been benchmarked on BEAM-100K, BEAM-1M, BEAM-10M, and LoCoMo10, showing leading performance/cost results under matched methodology.
Can I host AtomicMemory myself?
Yes, AtomicMemory is designed to be self-hostable. The core engine is open source and can be run locally or on your own infrastructure.
What is the difference between AtomicMemory and other memory products?
AtomicMemory prioritizes portability and inspectability, allowing the interface to be used across SDKs, CLI, MCP server, and adapters, with an open source engine that enables self-hosting and fact revision, unlike many hosted black-box solutions.