Honcho
Memory library for building stateful agents
Last verified:
What is Honcho?
Honcho is an AI-native open source memory library with a managed service for building stateful agents. It enables agents to build and maintain state about any entity including users, agents, groups, ideas, and more. As a continual learning system, Honcho understands entities that change over time by reasoning about stored messages to generate insights.
Key features include: automatic context management with smart conversation summaries for infinite chats, native multi-agent support with sessions that can have multiple participants, a powerful search endpoint combining full-text and semantic search using reciprocal rank fusion, the Dialectic API for natural language queries about entities, peer representations that provide explicit and deductive facts about users, and provider-agnostic design that works with any LLM or agent framework. Honcho uses formal logical reasoning performed by custom models to extract latent information from data.
Honcho is designed for developers building AI agents and LLM-powered applications that need personalized, stateful experiences. It's ideal for teams creating highly personalized experiences, agents with social cognition, agents with rich identity that evolve over time, and multi-agent systems with complex social dynamics. The SDK is available for Python and TypeScript.
Honcho pricing
Pricing model: Freemium
New tenants get $100 in free credits on sign up when creating an organization at app.honcho.dev. Honcho uses token-based pricing with tiers in v3 - it charges for the reasoning layer including ingestion and inference, while storage and retrieval are free. Unlike other memory frameworks that charge to access your own data, Honcho only charges for reasoning passes. The free tier is based on context ingested, not number of users, so you can have unlimited users limited only by content processing. Enterprise custom plans and a startup program offering $1,000 in credits with 12 months subsidized pricing are available.
Honcho pros
- Open source memory library with managed service option
- Perfect recall for agents with state-of-the-art long-term memory
- Automatic context management with smart conversation summaries
- Native multi-agent support with unlimited session participants
- Provider agnostic - works with any LLM or agent framework
- Powerful hybrid search combining full-text and semantic search
- Dialectic API enables natural language queries about entities
- Formal logical reasoning extracts latent information from data
- Custom models optimized for formal logical reasoning at lower cost
- Token batching reduces ingestion costs around 1000 tokens per batch
- $100 free credits for new tenants on sign up
- SDKs available for Python and TypeScript
- MCP connections and agent-first API design
- Peer cards provide key biographical information about users
- Context can be converted to OpenAI and Anthropic formats
- Asynchronous reasoning ensures fast writes without blocking
- SOTA benchmark performance on agent memory evaluations
- 5x cost reduction compared to previous versions in v3
Honcho cons
- Reasoning is asynchronous so insights are not immediately available
- Pricing based on reasoning passes (ingestion and inference) not transparent
- Queue status shows completed counts are not lifetime totals due to cleanup
- Summary mode required for very long conversations to control tokens
- Token batching means short messages wait for 1000 token threshold
- Demo server at demo.honcho.dev is not production ready
- Requires API key and account setup at app.honcho.dev to use
- Custom reasoning models still being actively improved for better quality
Frequently asked questions about Honcho
What is Honcho and what does it do?
Honcho is an AI-native open source memory library with a managed service for building stateful agents. It enables agents to build and maintain state about any entity including users, agents, groups, ideas, and more. As a continual learning system, it understands entities that change over time by storing messages and using formal logical reasoning to generate insights about each entity. Agents using Honcho have perfect recall and can get the exact context they need when they need it.
How does Honcho's reasoning work?
Honcho uses custom models trained specifically for formal logical reasoning. When you write messages to Honcho, they are stored immediately and enqueued for background processing. The reasoning models extract what was explicitly stated, draw certain conclusions from those premises, identify patterns across multiple conclusions, and infer the simplest explanations for behavior. These conclusions are stored as peer representations that can be queried for rich context. Reasoning is asynchronous and uses token batching around 1000 tokens to optimize costs.
What are the core data primitives in Honcho?
Honcho has four storage primitives: Workspaces are top-level containers that isolate different applications or environments; Peers are any entities that persist but change over time (users, agents, objects, and more); Sessions are interaction threads between peers with temporal boundaries; Messages are units of data that trigger reasoning including conversations, events, activity, and documents. Peers and sessions have a many-to-many relationship.
How do I get started with Honcho?
To get started, go to app.honcho.dev to sign up and get an API key. Every new tenant gets $100 in free credits. Install the SDK using uv add honcho-ai for Python or npm install @honcho-ai/sdk for TypeScript. Initialize the client with your workspace_id and api_key, create peers for users and assistants, create sessions, add messages to sessions, and then query for insights using methods like user.chat() to get synthesized information about entities.
What SDKs are available for Honcho?
Honcho provides SDKs for Python and TypeScript. The Python SDK can be installed with uv add honcho-ai or pip install honcho-ai. The TypeScript SDK can be installed with npm install @honcho-ai/sdk, yarn add @honcho-ai/sdk, or pnpm add @honcho-ai/sdk. Both SDKs provide client initialization, peer creation, session management, message addition, context retrieval, and queue status checking.
How does Honcho handle long conversations?
Honcho uses automatic summary generation at intervals to handle long conversations. The context() method includes a blend of summary and messages by default, covering the entire session history. You can specify any token limit you want, and can disable summaries to fill the limit entirely with recent messages using summary=False. For very long conversations, enable summary mode to maintain context while controlling token usage.
What is the Dialectic API?
The Dialectic API is an endpoint that lets you chat with Honcho about any entity in your system using natural language queries. Honcho leverages what it has remembered and learned about the entity to provide in-context actionable insights. Example queries include 'What's the best way to explain technical concepts to this user?', 'Is this user more task-oriented or relationship-oriented?', 'What time of day is this user most engaged?', and 'What are this user's core values based on our conversations?'
How does Honcho's search work?
Honcho implements a powerful search endpoint that searches for messages across a workspace, session, or peer with complex filters. The search combines full-text and semantic search using reciprocal rank fusion. By default, all messages have embeddings generated and stored, enabling semantic search. Results are returned as Message objects ordered by relevance, not recency. The default is 10 results with a maximum of 100. You can sort by timestamp or apply filters on created_at for time range searching.
What is queue status and why does reasoning take time?
Queue status shows the status of background reasoning tasks. When messages are stored, background processes kick off to reason about conversations and generate insights. Reasoning is asynchronous by design to efficiently process batches of messages rather than assessing each in isolation. Queue status tracks completed, in-progress, and pending work units. Task types include representation (memory formation), summary (session summarization), and dream (memory consolidation). There is a short processing delay before insights are available.
How does Honcho compare to traditional RAG?
Traditional RAG systems treat memory as static storage and retrieve what was explicitly said when semantically similar queries appear. They miss insights only accessible by reasoning about data, struggle with contradictions, and fail when predicting under uncertainty. Honcho extracts all latent information by reasoning about everything using formal logic, generating new understanding beyond simple recall. This produces more robust reasoning with exhaustive context for statefulness, earning higher retention and trust for agents.