# Honcho By **plastic-labs** ยท Software & Tools Open-source memory and context engine for stateful AI agents, works with any model, framework, or architecture. - Source: https://docs.honcho.dev - Repository: https://github.com/plastic-labs/honcho - Tags: open-source, memory, agents, stateful, rag, sdk, mcp, self-hosted - Pricing: freemium - Upvotes: 0 ## Features - Reasoning-first memory: extracts conclusions, not just retrieving chunks - Peer-centric model tracking users, agents, groups, and ideas as changing entities - Hybrid search (BM25 + vector) at peer, session, and workspace scope - Static low-latency representation endpoint for time-sensitive reads - Prompt-ready context bundles with .to_openai() and .to_anthropic() helpers - MCP server compatible with Claude Code, Cursor, Cline, Windsurf, and more - File upload for importing documents into peer context - Self-hostable FastAPI server with Docker; Python and TypeScript SDKs --- ## Why it matters Every agent builder eventually hits the same wall: RAG pipelines that retrieve what was said but miss what it *means*. Honcho breaks that cycle by adding a reasoning layer between raw message storage and context delivery, so agents understand users and agents as entities that change over time, not just bags of retrieved chunks. ## The big picture Honcho is built by Plastic Labs around a peer-centric model where humans and AI agents are first-class entities called peers. Peers participate in sessions, sessions hold messages, and a background reasoning system (powered by their custom Neuromancer models) continuously updates per-peer representations. The docs describe the architecture as four composable primitives: Workspaces, Peers, Sessions, and Messages. ## How it works The Honcho Loop has four steps per the README: store messages or events on a session, let Honcho reason asynchronously in the background, query for context (peer representations, session summaries, or natural-language answers via the Chat endpoint), then inject that context into any LLM call. Hybrid search (BM25 + vector) is available at the peer, session, and workspace level. For latency-sensitive reads, the static representation endpoint skips the reasoning queue entirely. ## By the numbers Per the Honcho evals page and README, the system claims 60-90% token savings over naive retrieval, with benchmark scores of 89.9% on LoCoMo, 90.4% on LongMem S, and BEAM scores ranging from 0.409 to 0.646 across context windows from 100K to 10M tokens. Open benchmarks are published on GitHub so results are verifiable. ## Zoom in Integrations cover the main agentic coding environments: Claude Code (via a marketplace plugin or raw MCP), OpenCode (npm plugin), OpenClaw (for messaging platforms including WhatsApp, Telegram, Discord, and Slack), Hermes, and any MCP-compatible client such as Cursor or Cline. For teams wiring Honcho into an existing codebase, the README describes an agent skill installed via `npx skills add plastic-labs/honcho` that explores your codebase and generates the SDK setup. ## Yes, but Background reasoning is asynchronous, so freshly added messages may not yet appear in chat or representation responses. The README flags this explicitly and recommends the low-latency representation endpoint for time-sensitive reads. Teams that need deterministic, synchronous retrieval should factor this into their latency budget. ## The bottom line If you have moved past simple RAG and need agents that track how users and other agents change over time, Honcho is the most complete open-source primitive for that job today. The managed service removes infrastructure overhead; the self-hosted path keeps data in-house. Either way, the reasoning layer is what separates it from a vector store with a fancy API. ## FAQ ### What is Honcho and what problem does it solve? Honcho is open-source memory infrastructure for stateful AI agents, built by Plastic Labs. It solves the core limitation of RAG-based memory: retrieval systems can surface what was said, but they miss implicit patterns, contradictions, and predictions that only emerge from reasoning. Honcho stores messages and events, reasons over them in the background using custom Neuromancer models, and returns structured peer representations that any agent framework can consume. ### How do I install and start using Honcho? Install the Python SDK with `pip install honcho-ai` (or `uv add honcho-ai`) or the TypeScript SDK with `npm install @honcho-ai/sdk`. Get an API key by signing up at app.honcho.dev, which also sets up an organization with $100 in free credits. For the managed service, pass your API key and a workspace ID; for self-hosting, point the SDK at `http://localhost:8000` after running the FastAPI server via Docker. The README includes a full working example covering store, reason, query, and inject in roughly 20 lines of Python or TypeScript. ### Is Honcho free or open source? The core Honcho server is open source and available on GitHub under the plastic-labs organization. The managed service at api.honcho.dev gives new organizations $100 in free credits on signup; costs beyond that are not detailed in the public sources. Self-hosting is fully supported via Docker, so you can run the entire system without the managed service. ### What is Honcho best for? Honcho is best for products where user continuity and personalization across many sessions is a core requirement, such as tutoring agents, coaching tools, or any assistant that needs to remember how a specific user thinks and changes over time. The peer-centric model also makes it a strong fit for multi-agent systems where agents need to model what other agents know. Per the docs, it is designed as a general solution to context engineering, memory, and statefulness rather than a narrow point solution. ### How does Honcho compare to a standard RAG or vector store setup? Standard RAG retrieves stored text chunks that match a query; Honcho reasons over stored messages to extract conclusions, track patterns, handle contradictions, and make predictions, then surfaces those as queryable peer representations. Per the README, this approach claims 60-90% token savings because Honcho delivers curated context rather than raw retrieved chunks. The tradeoff is that reasoning is asynchronous, so there is a brief lag between writing messages and seeing updated representations, whereas a vector store retrieval is synchronous. ### What are the main limitations or risks of using Honcho? The most important operational limitation is that background reasoning is asynchronous: messages written to Honcho may not yet be reflected in chat or representation queries for a short period, which the README explicitly flags. Architectures that require synchronous, deterministic retrieval need to use the low-latency static representation endpoint and design around this delay. On the managed service, data is routed through api.honcho.dev and pricing beyond the initial $100 credit is not publicly documented, so cost at scale requires direct engagement with the Plastic Labs team. --- [View on Analog](https://analoghq.ai/plastic-labs/software-tools/honcho)