plastic-labs/honcho

What it solves

Honcho provides memory infrastructure for stateful AI agents, allowing them to maintain a persistent understanding of people, other agents, groups, projects, and ideas as they evolve over time. It moves beyond simple chunk-matching (standard RAG) to a reasoning-first approach that extracts conclusions from interactions to build evolving representations of entities.

How it works

Honcho operates through a continuous loop of storing, reasoning, and querying:

1. Store: Conversations, events, documents, or tool traces are saved as messages within a session.
2. Reason: In the background, Honcho asynchronously processes these messages to update "peer representations" (what the system knows about a specific entity).
3. Query: Users can query these representations via natural language, search for specific results, or pull prompt-ready context for an LLM.
4. Inject: The resulting context or answer is injected into an LLM call or agent framework.

Internally, it uses a peer-centric model where humans and agents are treated as first-class entities (peers) who participate in sessions. It stores observations in collections of vector-embedded documents keyed by observer-observed peer pairs.

Who it’s for

• Developers building AI agents that require long-term memory and high retention.
• Teams creating multi-agent systems where agents need to understand each other and their users.
• Users of coding agents (like Claude Code, Cursor, or OpenCode) who want a shared, persistent memory across different tools.

Highlights

• Reasoning-first memory: Extracts actual conclusions from conversations rather than just retrieving similar text chunks.
• Peer-centric model: Tracks the evolution of users, agents, and projects as distinct entities.
• Multi-peer perspective: Can model what one specific peer knows about another.
• Broad Integration: First-party plugins for major coding agents including Claude Code, Cursor, and DeepSeek Harness, as well as MCP (Model Context Protocol) support.
• Flexible Deployment: Available as a managed service, a local stack via CLI, or a self-hosted FastAPI server.