atomicmemory
starsforks
How to install atomicmemory?
atomicstrata/atomicmemory
$git clone https://github.com/atomicstrata/atomicmemoryInspectable, portable semantic memory for agents and applications.
README.md
| 1 | # AtomicMemory |
| 2 | |
| 3 | [](https://github.com/atomicstrata/atomicmemory/actions/workflows/ci.yml) |
| 4 | [](https://www.npmjs.com/package/@atomicmemory/core) |
| 5 | [](https://www.npmjs.com/package/@atomicmemory/sdk) |
| 6 | [](https://www.npmjs.com/package/@atomicmemory/cli) |
| 7 | [](packages/core/Dockerfile) |
| 8 | [](https://docs.atomicstrata.ai) |
| 9 | [](LICENSE) |
| 10 | |
| 11 | **Inspectable, portable semantic memory for agents and applications.** |
| 12 | |
| 13 | AtomicMemory is a memory layer you embed where your AI code already runs. Capture |
| 14 | context, ground generations in prior interactions, and carry knowledge across |
| 15 | sessions — from a direct SDK call, a CLI, an MCP server, a framework adapter, or |
| 16 | a host plugin. Local-first where supported, hosted where convenient, and |
| 17 | designed so the choice can change later without rewriting your application. |
| 18 | |
| 19 | Most memory products ask you to trust a hosted black box with the layer that |
| 20 | decides what an AI believes about your users. AtomicMemory takes the opposite |
| 21 | position: the interface should be portable, the engine should be inspectable, |
| 22 | and the memory system should be able to revise itself when facts change. |
| 23 | |
| 24 | This repository is the public source of truth for the AtomicMemory JavaScript / |
| 25 | TypeScript packages, framework adapters, host plugins, and public smoke tests. |
| 26 | |
| 27 | **Docs:** [docs.atomicstrata.ai](https://docs.atomicstrata.ai) |
| 28 | |
| 29 | **Field note:** [The AI Memory Industry Has A Black Box Problem](https://www.atomicstrata.ai/blog/the-ai-memory-industry-has-a-black-box-problem) |
| 30 | |
| 31 | ## Headline Results |
| 32 | |
| 33 | AtomicMemory v66 is leading performance/cost on BEAM-100K, BEAM-1M, and LoCoMo10 under |
| 34 | matched methodology against published competitors. On BEAM-10M it matches the |
| 35 | strongest published Mem0-new result while leaving Hindsight-scale temporal |
| 36 | retrieval as the known open frontier. |
| 37 | |
| 38 | | Benchmark | AtomicMemory v66 | Position | Cost/Q | Sample | |
| 39 | |---|---:|---|---:|---:| |
| 40 | | **BEAM-100K lenient** | **0.7375** | Parity with Hindsight at 0.75 | $1.26 | n=80 | |
| 41 | | **BEAM-1M lenient** | **0.6625** | Leading Performance/Cost; +0.022 vs Mem0 paper | $0.083 | n=80 | |
| 42 | | **BEAM-10M lenient** | **0.4875** | Parity with Mem0-new at 0.486 | $0.081 | n=80 | |
| 43 | | **LoCoMo10 GPT-4o-mini binary** | **0.8396** | Leading Performance/Cost; +0.171 vs Mem0 paper | $0.066 | n=1540 | |
| 44 | |
| 45 | These results put AtomicMemory at or near the published ceiling in each |
| 46 | reported category while preserving the lower-cost operating profile that |
| 47 | matters for real applications. Reproducibility artifacts and harness details |
| 48 | will be published with the benchmark materials. |
| 49 | |
| 50 | ## Why AtomicMemory |
| 51 | |
| 52 | - **Portable**: a single memory protocol consumed by direct SDK calls, CLIs, |
| 53 | the MCP server, framework adapters, and host plugins. The same memory store |
| 54 | serves a LangGraph agent, a Claude Code session, and a custom Vercel AI |
| 55 | application without re-implementing capture or retrieval semantics. |
| 56 | - **SDK-agnostic**: every adapter is built on the same SDK. Adapters are |
| 57 | conveniences, not gatekeepers. You can drop down to the SDK at any time and |
| 58 | keep the same data, indexes, and retrieval behavior. |
| 59 | - **Inspectable**: Core is open source, self-hostable, and built around |
| 60 | explicit mutation decisions rather than an opaque hosted opinion. |
| 61 | - **Correction-aware**: memory is not just append and recall. Real products |
| 62 | need supersession, clarification, deletion, no-op decisions, lineage, and |
| 63 | trust-sensitive revision when users change their mind. |
| 64 | - **Model-surface portable**: the SDK lets applications swap memory backe |
People who liked this also viewed
6 picksType
- excel-mcpA Model Context Protocol (MCP) server for working with Excel (.xlsx) files without requiring Microsoft Excel. Enables AI assistants to create, read, modify, and format spreadsheets using openpyxl, fully in Python and cross-platform.MCP ServersOct 2025
- memories-with-lessons-mcp-serverI enhance the existing memory mcp server from the official mcp github, so big thanks and credits for creating thisMCP ServersFeb 2025
- anytype-mcpAn MCP server enabling AI assistants to interact with Anytype - your encrypted, local and collaborative wiki - to organize objects, lists, and more through natural language.MCP ServersJul 2026
- mcp-server-kubernetesMCP Server for kubernetes management commandsMCP ServersJul 2026
- thoughtboxThoughtbox is an intention ledger for agents. Evaluate AI's decisions against its decision-making.MCP ServersJul 2026
- google-drive-mcpA Model Context Protocol (MCP) server that provides secure integration with Google Drive, Docs, Sheets, Slides and Calendar. It allows Claude Desktop and other MCP clients to manage files in Google Drive through a standardized interface.MCP ServersJul 2026
