# Mnemo Memory

> Long-term memory infrastructure for AI agents. Hybrid retrieval (vector + keyword + knowledge graph + temporal + reranking). Benchmarked publicly on LongMemEval-S: 85.2% on the official LongMemEval judge. US-hosted with US data residency. Free tier: 1,000 memory writes + 10,000 searches per month. No credit card.

Mnemo is a drop-in memory API for AI agents. You write facts; we store them with atomic fact extraction. You query in natural language; we run multiple retrieval strategies in parallel and fuse the results. Unlike vector-only stores, we combine semantic similarity, keyword matching, knowledge graph edges, temporal decay, and reranking — this is why we score 85.2% on the official LongMemEval judge (gpt-4o-mini) on LongMemEval-S while pure-vector approaches plateau lower.

Mnemo is independent of your agent's generation model. Keep using your existing model and framework while Mnemo handles memory storage and retrieval. Self-hosting is on the Enterprise roadmap.

Every page on mnemohq.com also serves a Markdown representation of itself: send `Accept: text/markdown` to any page URL and you get the prose without the layout.

## When to use Mnemo

- [Long-term memory for AI agents](https://mnemohq.com/for-agents): the agent must recall what a user said days or weeks ago, not just what fits in the current context window. Write with POST /v1/memories, read back with POST /v1/search.
- [Cited answers over ingested documents](https://mnemohq.com/docs#api-documents): you need an answer that carries its source. Every Mnemo result returns the memory, its score, and the document it came from — a citation, not a vibe.
- [Cross-session user preference recall](https://mnemohq.com/docs#api-profile): the agent should stop re-asking things the user already told it. POST /v1/profile returns a prompt-ready block of what is known about one user.
- [Team memory inside Slack](https://mnemohq.com/brain): Company Brain learns from the public channels you invite it to and answers with links back to the exact messages.
- [Memory inside an existing agent framework](https://mnemohq.com/integrations): LangChain, LangGraph, Vercel AI SDK, AutoGen, CrewAI, PydanticAI, Semantic Kernel, LlamaIndex, Mastra, or MCP — Mnemo slots in beside the framework rather than replacing it.
- [Answers on your own site](https://mnemohq.com/widget): one script tag turns your docs into a question box that answers with numbered sources.

## When not to use Mnemo

- Not a vector-database primitive. If you want to own chunking, embedding models, index parameters, and raw ANN queries, use pgvector, Qdrant, or Pinecone directly. Mnemo makes those decisions for you on purpose.
- Not a general search engine. Mnemo only ever returns what was written into your workspace. It does not crawl the web and has no world knowledge of its own.
- Not a document store or CMS. Retrieval returns atomic facts with provenance, not the original files back. Keep the files where they already live.
- Not a system of record. Mnemo is durable and event-sourced, but your primary database should stay primary — treat Mnemo as a derived, rebuildable index of what matters.
- Not a chat model. Mnemo does not generate the reply; it supplies the context your model generates from.

## Developer resources

- [Mnemo API documentation](https://mnemohq.com/docs): quickstart, full REST reference, SDK guide, integrations, versioning policy, and self-hosting. Base URL https://api.mnemohq.com.
- [Mnemo OpenAPI 3.1 spec](https://mnemohq.com/openapi.json): the machine-readable contract for every endpoint. Import it into a client generator or an agent tool loader.
- [Mnemo CLI](https://www.npmjs.com/package/getmnemo-cli): `npm install -g getmnemo-cli` — add, search, and inspect memories from a terminal or a CI job.
- [Mnemo TypeScript SDK](https://www.npmjs.com/package/getmnemo): `npm install getmnemo` — zero runtime dependencies, works on Node 18+, Bun, Deno, browsers, and Cloudflare Workers.
- [Mnemo Python SDK](https://pypi.org/project/getmnemo/): `pip install getmnemo`.
- [Mnemo MCP server](https://www.npmjs.com/package/getmnemo-mcp): `npx -y getmnemo-mcp` for a local server, or connect to the hosted endpoint at https://mcp.mnemohq.com/mcp. Exposes search, add, get, update, delete, and list as MCP tools for Claude Desktop, Cursor, Windsurf, VS Code, and Zed.
- [Mnemo LangGraph package](https://www.npmjs.com/package/getmnemo-langgraph): `npm install getmnemo-langgraph` — recall and store nodes for a LangGraph state graph.
- [Mnemo Vercel AI SDK package](https://www.npmjs.com/package/getmnemo-vercel-ai): `npm install getmnemo-vercel-ai` — typed tools for streamText, generateText, and useChat.
- [Mnemo Anthropic SDK package](https://www.npmjs.com/package/getmnemo-anthropic): `npm install getmnemo-anthropic` — Mnemo as a memory tool with prompt caching and extended thinking wired up.
- [Mnemo Mastra package](https://www.npmjs.com/package/getmnemo-mastra): `npm install getmnemo-mastra` — implements Mastra's memory provider interface.
- [Mnemo dashboard](https://app.mnemohq.com): create a workspace, mint and rotate API keys, watch usage, configure connectors and widget tokens.
- [Mnemo API versioning and deprecation policy](https://mnemohq.com/docs#api-versioning): v1 is URL-versioned; breaking changes only ever ship in a new version path; deprecations are announced in the changelog at least six months before sunset, with `Deprecation` and `Sunset` headers on every affected response.

## Authentication

- [Get a free Mnemo API key](https://app.mnemohq.com/signup?utm_source=marketing&utm_medium=site_cta): 1K memory writes + 10K searches/month, no credit card.
- [Authentication reference](https://mnemohq.com/docs#auth): every request sends `Authorization: Bearer <api key>` plus `x-workspace-id: <workspace>`. Keys are workspace-scoped, hashed at rest, revocable, and rotatable with a 24-hour grace window. There is no OAuth flow and no anonymous tier — issue keys from the dashboard.

## Getting Started

- [Quickstart](https://mnemohq.com/docs#quickstart): 5-minute setup with TypeScript, Python, or cURL
- [Live playground](https://mnemohq.com/playground): Watch the retrieval strategies race in real time
- [For AI agents](https://mnemohq.com/for-agents): Copy-paste tool schemas, MCP install, IDE rules
- [Pricing](https://mnemohq.com/pricing): Free, Pro, Team, Enterprise tiers
- [Contact](https://mnemohq.com/contact): support, sales, security disclosure, and partnership channels

## API Reference

- [OpenAPI 3.1 spec](https://mnemohq.com/openapi.json): Machine-readable full API specification
- [POST /v1/memories](https://mnemohq.com/docs#api-memories): Write a memory; returns extracted atomic facts
- [POST /v1/documents](https://mnemohq.com/docs#api-documents): Async ingestion of transcripts or long docs
- [POST /v1/search](https://mnemohq.com/docs#api-search): Hybrid retrieval across all strategies, fused, with citations
- [POST /v1/profile](https://mnemohq.com/docs#api-profile): Prompt-ready context block for the current user
- [Errors and rate limits](https://mnemohq.com/docs#api-errors): Uniform error bodies, Retry-After and X-RateLimit-* headers

## Integrations

- [All integrations](https://mnemohq.com/integrations): 8 frameworks with code examples
- [LangChain](https://mnemohq.com/integrations/langchain): Decorated tool functions (Python)
- [LangGraph](https://mnemohq.com/integrations/langgraph): Graph nodes with recall + store
- [Vercel AI SDK](https://mnemohq.com/integrations/vercel-ai-sdk): Typed Zod tool definitions (TypeScript)
- [AutoGen](https://mnemohq.com/integrations/autogen): Multi-agent shared memory
- [CrewAI](https://mnemohq.com/integrations/crewai): Crew-level memory tools
- [PydanticAI](https://mnemohq.com/integrations/pydantic-ai): Type-safe agents with RunContext
- [Semantic Kernel](https://mnemohq.com/integrations/semantic-kernel): C# kernel plugin
- [LlamaIndex](https://mnemohq.com/integrations/llamaindex): RAG + user memory hybrid
- [MCP server](https://mnemohq.com/for-agents#mcp): Install in Claude Desktop, Cursor, Windsurf, VS Code, Zed
- [IDE rules](https://mnemohq.com/for-agents#ide-rules): .cursorrules, CLAUDE.md, .windsurfrules, AGENTS.md
- [Data-source connectors](https://mnemohq.com/connectors): 20 sources — Google Drive, Gmail, Notion, Slack, HubSpot, GitHub, Linear, and more — with backfill, incremental sync, and deletion propagation

## Why Mnemo

- [Benchmarks](https://mnemohq.com/benchmarks): Full per-category LongMemEval breakdown — no cherry-picking
- [Leaderboard](https://mnemohq.com/benchmarks/leaderboard): Sortable per-category table across memory systems
- [Product](https://mnemohq.com/product): How the hybrid retrieval pipeline works
- [ROI calculator](https://mnemohq.com/roi): Compare $/mo vs rolling your own pgvector + Redis + reranker stack
- [Migration wizard](https://mnemohq.com/migrate): Paste your existing memory-API config, get the Mnemo equivalent

## Trust & Security

- [Trust center](https://mnemohq.com/trust): Security covenants, compliance, data handling
- [Customers](https://mnemohq.com/customers): Case studies and logos
- [About](https://mnemohq.com/about): Mission and roadmap
- [Privacy policy](https://mnemohq.com/privacy): What we collect, sub-processors, retention, your rights
- [Terms of service](https://mnemohq.com/terms): The agreement covering API use

## Optional

- [Blog](https://mnemohq.com/blog): Engineering deep-dives on retrieval architecture
- [Changelog](https://mnemohq.com/changelog): Release notes, and where deprecations are announced
- [Sitemap](https://mnemohq.com/sitemap.xml): Every URL on this site
- [llms-full.txt](https://mnemohq.com/llms-full.txt): Single-file concatenated docs for one-shot LLM ingestion