Honcho

Open-source memory layer that reasons about every user your agent meets

Advanced API
Visit HonchoBack to AI tools

Own Honcho? Verify ownership to get a dofollow link

Screenshot of Honcho, Open-source memory layer that reasons about every user your agent meets

What is Honcho?

Honcho is open-source memory infrastructure for AI agents. Messages written to it trigger reasoning that produces compact, curated context, so agents remember users across sessions and tools while spending far fewer tokens. It publishes benchmark results on LongMem, LoCoMo and BEAM and targets developers building stateful agents.

Honcho is an open-source memory layer for AI agents, built by Plastic Labs. Instead of treating memory as a store-and-retrieve problem, it runs reasoning over every message an agent ingests, producing curated context rather than a pile of retrieved facts. The pitch is twin-fold: agents keep continuity with a given user across sessions, tools and machines, and the tokens spent on context drop sharply because the system delivers a distilled summary plus relevant history instead of re-reading a long transcript. The mechanics are deliberately small. A developer writes messages into Honcho, where they are stored and indexed, and a custom reasoning model called Neuromancer draws conclusions in the background. Calling context() then returns instant curated reasoning plus conversation history, with knobs for search, token budgets, summaries and peer scoping. A separate .chat() call handles targeted questions at five effort tiers, from a single semantic lookup at the minimal end up to asynchronous multi-pass analysis at the top. The company reports 60 to 90 percent token savings and publishes benchmark results on LongMem, LoCoMo and BEAM. The model departs from the usual user-assistant framing. Peers can represent people, agents, NPCs or whole groups, and their perspectives can be scoped globally or narrowly across sessions. A background process called dreaming keeps refining each peer's profile without affecting runtime latency. That flexibility points at companions, coding agents, game characters, tutoring systems and support desks, though the primary audience is developers comfortable wiring an SDK into their own stack. It competes with memory products such as Mem0, and its differentiation rests on reasoning quality and openness rather than a polished end-user app.

How do you use Honcho?

  1. 1Read the quickstart
    Work through the five-minute quickstart to see how messages, sessions and peers fit together before writing any code.
  2. 2Create a workspace and get an API key
    Sign in to the app, set up a workspace, and copy the credentials your SDK calls will use.
  3. 3Ingest messages for a peer
    Send conversation messages through the SDK or CLI so they are stored, indexed and picked up by Neuromancer for reasoning.
  4. 4Pull context into your agent
    Call context() each turn to receive the curated summary plus history, and set a token budget to control how much comes back.
  5. 5Query at the right reasoning tier
    Use .chat() for targeted questions and pick an effort level between minimal and max depending on how deep the answer needs to go.
  6. 6Connect your coding agent
    Follow the vibecoding setup to attach Honcho to tools such as Claude Code, Codex or OpenClaw so memory persists across sessions.

Pros and cons

Pros

  • Reasoning-based memory goes past retrieval, producing conclusions and patterns rather than raw fact lookupAI
  • context() responds in roughly 200ms, fast enough to call on every conversation turnAI
  • Published 60 to 90 percent token savings and results across LongMem, LoCoMo and BEAMAI
  • Peer modeling extends beyond user and assistant to agents, NPCs, groups and scoped perspectivesAI
  • Open-source core with a free tier and a startup credit program for small teamsAI

Cons

  • Reasoning calls are metered in five tiers, so heavy use of high or max queries adds up quickly against the stated per-query pricesAI
  • The product is infrastructure: there is no polished end-user application, and teams must integrate the SDK themselvesAI
  • Benchmark leadership is self-published, and buyers should verify results against their own workloads before committingAI
  • Background dreaming quality and its long-term effect on context are hard to evaluate up frontAI
  • Documentation is versioned and still evolving, which can slow teams unfamiliar with the primitivesAI

How much does Honcho cost?

Pricing

Memory ingestion with Neuromancer reasoning costs $2.00 per million tokens, context() calls are unlimited at about 200ms, and standard dreaming is included. Reasoning queries are priced per call by tier: $0.001 minimal, $0.01 low, $0.05 medium, $0.10 high and $0.50 max. Startups under $5M raised can apply for $1,000 in credits and 12 months of subsidized pricing; enterprise plans are custom.

Learn more

Support

Support runs through a Discord community, full SDK documentation with guides, and a published GitHub repository for the open-source core. Enterprise buyers get forward-deployed engineers plus dedicated integration and maintenance support.

Learn more

Integrations

An agent skill is listed for Claude Code, Codex, Hermes Agent, OpenClaw, Kilo Code, Pi, OpenCode and DeepSeek Harness, with a dedicated vibecoding setup guide. The reasoning layer is model-agnostic across OpenAI, Anthropic and custom models.

Learn more

Features

Messages written to Honcho are stored, indexed and reasoned over automatically by Neuromancer. context() returns curated reasoning plus history in about 200ms; .chat() handles targeted queries across five effort tiers. Peers model users, agents, NPCs and groups with scoped perspectives, and dreaming refines each profile in the background without touching runtime latency.

Learn more

Frequently asked questions about Honcho

  • How much does Honcho cost?
    Memory ingestion with Neuromancer reasoning costs $2.00 per million tokens, context() calls are unlimited at about 200ms, and standard dreaming is included. Reasoning queries are priced per call by tier: $0.001 minimal, $0.01 low, $0.05 medium, $0.10 high and $0.50 max. Startups under $5M raised can apply for $1,000 in credits and 12 months of subsidized pricing; enterprise plans are custom.
  • Does Honcho have an API?
    Yes, Honcho offers an API.
  • How do you use Honcho?
    The walkthrough on this page covers 6 steps: 1. Read the quickstart 2. Create a workspace and get an API key 3. Ingest messages for a peer 4. Pull context into your agent 5. Query at the right reasoning tier 6. Connect your coding agent.
  • What platforms does Honcho support?
    Honcho is available on Web App, Linux, MacOS and Windows.
  • What does Honcho integrate with?
    An agent skill is listed for Claude Code, Codex, Hermes Agent, OpenClaw, Kilo Code, Pi, OpenCode and DeepSeek Harness, with a dedicated vibecoding setup guide. The reasoning layer is model-agnostic across OpenAI, Anthropic and custom models.
  • What are the limitations of Honcho?
    Reasoning calls are metered in five tiers, so heavy use of high or max queries adds up quickly against the stated per-query prices. The product is infrastructure: there is no polished end-user application, and teams must integrate the SDK themselves. Benchmark leadership is self-published, and buyers should verify results against their own workloads before committing.

Status

StatusActive
Views0
Outbound clicks0
Added10/21/2026

Platforms

Web AppLinuxMacOSWindows

Pricing

FreemiumPaidSubscription

Categories

DeveloperPhonk Song GeneratorCodingMusicProductivity