What is Honcho?
Honcho is a memory server from Plastic Labs built around peers: users, agents, groups or projects that take part in sessions. You store messages on a session; a background worker called the deriver reads them and writes conclusions about each peer, both stated facts and inferred ones, and a separate dreaming process consolidates them later. You then ask for prompt-ready context, run a hybrid BM25 and vector search, or put a natural-language question to the chat endpoint, which reasons over what Honcho has concluded.
Conclusions are stored as vector-embedded documents in collections keyed by observer and observed peer, so Honcho can hold what one agent knows about another as well as each peer's own picture. Storage is PostgreSQL with pgvector by default, with turbopuffer, LanceDB and Qdrant as alternatives.
The FastAPI server is AGPL-3.0 and self-hostable with Docker Compose or the honcho start command; Plastic Labs also runs a managed service at api.honcho.dev.
Is Honcho worth it?
Honcho answers a different question from most memory layers. Mem0 and Zep return facts that match a query; Honcho keeps a running model of each participant and will answer a question about them in prose, such as what learning style a student responds to. That suits tutors, companions and support agents more than knowledge lookup.
The work happens after the write. A message is stored at once, then queued for the deriver, which extracts conclusions with an LLM (gpt-5.4-mini by default) and a later dreaming pass consolidates them. The README warns that new messages take a moment to show up in chat answers, so an agent that needs a fact from the previous turn should read the stored session rather than rely on reasoning that has not run yet.
Plastic Labs reports 90.4% on LongMemEval S and 89.9% on LoCoMo, against 62.6% and 75.6% for Claude Haiku 4.5 with the full conversation in context, and 92.6% on LongMemEval with Gemini 3 Pro answering. Those runs used gemini-2.5-flash-lite for extraction and claude-haiku-4.5 for reasoning, and the company notes its hosted service uses a variety of models. On BEAM the score falls from 0.630 at 100,000 tokens to 0.409 at 10 million, the clearest public sign of how any of these systems degrade with very long histories. LoCoMo was graded by an LLM judge rather than the paper's F1 metric, so it does not line up with older published scores.
The managed service is the cheapest hosted write among the memory layers we review, at $2.00 per million tokens ingested, with context reads unlimited at about 200 ms. Chat queries cost $0.01 at the default level and $0.50 at the maximum. A self-hosted server has auth off by default (AUTH_USE_AUTH=false) and needs a JWT secret before it goes on a network; telemetry and Sentry are off in the shipped template.
AGPL-3.0 is the catch for commercial teams: running a modified Honcho as a network service obliges you to publish your changes. Using it unmodified, or through the managed API, does not. The project is active, with v3.5.0 released on 30 October 2026 and about 8,000 GitHub stars.
Where it earns its place
An edtech start-up running an AI tutor for 10,000 students
Adapt explanations to how each student learns
Store each lesson as a session, then ask peer.chat what a student struggles with before the next one. At $2.00 per million tokens ingested, a student producing 50,000 tokens a month costs $0.10 to remember, plus a cent per default chat query.
A developer building a multi-agent game with characters that know different things
Keep each character's knowledge of the player and of each other separate
Configure which peers observe which; Honcho keeps a collection per observer and observed pair, so one character can be ignorant of what another has seen. Self-hosting suits this, with Postgres and one LLM key.
A solo developer using Claude Code, Cursor and Codex on the same projects
Give all three agents one shared memory of preferences and past decisions
honcho init writes one key for every plug-in, and each agent reads the same workspace. The $5 monthly credit with a card on file covers light personal use on the managed service.
Before you commit
- Setup
- uv tool install honcho-cli then honcho start --setup runs API, deriver, Postgres and Redis in Docker with one LLM key. From source: Docker Compose, or Python 3.13+, Postgres with pgvector, the API and a separate deriver process.
- Integrations
- Python and TypeScript SDKs, a CLI and an MCP server; first-party plug-ins for Claude Code, Codex, Cursor, OpenCode, Kilo Code, OpenClaw and Hermes. Any OpenAI-compatible LLM provider, including OpenRouter, vLLM and Ollama, as long as it supports tool calling.
- Data and privacy
- Self-hosted data stays in your Postgres or chosen vector store, but message text goes to the configured LLM. Telemetry and Sentry are off in the shipped template. The managed service gives each organisation its own instance; retention terms, training use and certifications are not published on the site.
- Cost at scale
- Managed: $2.00 per million tokens ingested, unlimited context reads, and $0.001 to $0.50 per chat query depending on reasoning level. Self-hosted: Postgres plus deriver and dreaming LLM calls on every processed message.
- Limits
- Reasoning is asynchronous, so conclusions lag recent messages. BEAM scores drop from 0.630 to 0.409 between 100,000 and 10 million tokens. 180 open issues and pull requests.
What’s good
- Builds a model of each person and agent, not just a fact list
- Cheapest hosted ingestion of the memory layers we review, at $2.00 per million tokens
- Tracks what one agent knows about another
- Telemetry off by default when self-hosted
What’s not
- AGPL-3.0 obliges you to publish changes to a hosted fork
- Background reasoning lags new messages
- Self-hosted auth is off by default
- No SOC 2 or HIPAA statement on the site
Best for
Teams building tutors, companions or support agents that must understand individual users over many sessions.
Honcho alternatives
Honcho: common questions
- Is Honcho worth it?
- Honcho is the memory layer to choose when you want an agent to understand a person rather than look facts up, and its AGPL licence is the first thing your lawyers will ask about. We rate it 4.1 out of 5.
- How much does Honcho cost?
- Server free under AGPL-3.0. Managed service: $2.00 per million tokens ingested (context reads unlimited, dreaming included); chat queries from $0.001 (minimal) to $0.50 (max) each, $0.01 by default. $5 in credits for new accounts and $5 a month once a card is added; Enterprise by quote.
- What are the downsides of Honcho?
- AGPL-3.0 obliges you to publish changes to a hosted fork; background reasoning lags new messages; self-hosted auth is off by default; no SOC 2 or HIPAA statement on the site.
- Who is Honcho best for?
- Teams building tutors, companions or support agents that must understand individual users over many sessions.
- How much does Honcho cost for a team or in production?
- Managed: $2.00 per million tokens ingested, unlimited context reads, and $0.001 to $0.50 per chat query depending on reasoning level. Self-hosted: Postgres plus deriver and dreaming LLM calls on every processed message.
- Is Honcho safe for company data?
- Self-hosted data stays in your Postgres or chosen vector store, but message text goes to the configured LLM. Telemetry and Sentry are off in the shipped template. The managed service gives each organisation its own instance; retention terms, training use and certifications are not published on the site.
- What are the best Honcho alternatives?
- The closest alternatives we have reviewed are Hindsight, Mem0, Graphiti and Letta.
Answers checked Nov 6, 2026.
Make Honcho? Show this rating on your site
