Agents & Multi-Agent

Honcho

plastic-labs/honcho

Memory infrastructure for building stateful AI agents that understand changing people, agents, groups, projects, and ideas over time.

Open official repository ↗ Project homepage ↗
★ 6.5KStars
⑂ 788Forks
179Open issues
PythonLanguage
AGPL-3.0License
Q@project.QualityScoreEditorial score

Project screenshots

Screenshot of Honcho Screenshot of Honcho Screenshot of Honcho Screenshot of Honcho Screenshot of Honcho Screenshot of Honcho

Overview

Honcho provides a reasoning-first memory system for AI agents. It processes conversations and events asynchronously in the background to extract conclusions and build dynamic representations of peers (users and agents). Developers can store messages, let Honcho reason over them, and then query for prompt-ready context, peer representations, or natural-language insights. It can be used as a managed service or self-hosted via a FastAPI server.

Key features

  • Reasoning-first memory extracting conclusions rather than just matching chunks
  • Peer-centric model tracking users, agents, and groups as entities that change over time
  • Multi-peer perspective modeling what one peer knows about another
  • Managed service (api.honcho.dev) or self-hosted FastAPI server
  • Agent-tool integrations via MCP, Claude Code, OpenCode, OpenClaw, and Hermes
  • Python and TypeScript SDKs
  • Hybrid search (BM25 + vector)
  • Asynchronous background processing for representations and summaries
  • Prompt-ready context generation for OpenAI and Anthropic APIs

Requirements, installation and quick start

Python SDK: `pip install honcho-ai`. TypeScript SDK: `npm install @honcho-ai/sdk`. Self-hosting via Docker: `git clone https://github.com/plastic-labs/honcho.git`, `cd honcho`, `cp docker-compose.yml.example docker-compose.yml`, `cp .env.template .env` (fill in LLM API keys), `docker compose up`.

Usage

Initialize the Honcho client with a workspace ID and API key. Create peers and a session, then add messages to the session. Background reasoning happens automatically. Query peer insights using `peer.chat()` or retrieve prompt-ready session context using `session.context()`. Inject the context into an LLM call using format converters like `context.to_openai()` or `context.to_anthropic()`.

Model compatibility and use cases

Configurable multi-provider LLM support. Google Gemini is used by default for deriver, summary, and dialectic minimal/low tasks. Anthropic is used by default for dialectic medium/high/max and dream tasks. OpenAI is used for embeddings when EMBED_MESSAGES=true.

License and risk notes

Repository metadata declares AGPL-3.0. Before publishing or commercial use, review the repository LICENSE file, copyright notices and dependency terms.

Editorial verification 2026-08-09: repository URL, owner, description, license and repository statistics were reviewed. License metadata: AGPL-3.0. README was fetched for the channel draft; re-check repository dependencies, releases and model terms before production use.

Release and maintenance

Server version 3.0.9

Firecrawl

firecrawl/firecrawl

★ 161.1KTypeScript

LangChain

langchain-ai/langchain

★ 143.6KPython

RAGFlow

infiniflow/ragflow

★ 86.7KGo