Honcho: Open-Source AI Memory That Reasons, Explained
Honcho is an open-source memory system for AI agents, made by Plastic Labs. Every time an agent sends Honcho a message, Honcho stores it and has its own reasoning models work out what the message says about the user: facts, but also conclusions, patterns and preferences. When the agent needs context, one call returns a curated summary of what matters plus recent conversation history. It works with any model or framework, can be self-hosted, and has plugins that give Claude Code, Codex, OpenClaw and Hermes Agent memory across sessions and machines. Its GitHub repository has about 8,100 stars.
The quick answer: Honcho’s distinguishing idea is that memory should be reasoning, not storage: rather than filing away facts, it draws conclusions about each person and keeps refining them in the background. It’s popular with developers building companions and coding agents, especially in the Hermes Agent community. A disclosure: we make Carly, an AI assistant with its own memory and workspace, so we compete on the memory side. Everything below comes from Honcho’s site, docs and funding coverage.
Honcho at a glance
Checked against Honcho’s site, documentation and funding coverage on November 12, 2026.
| Honcho | |
|---|---|
| What it is | Open-source memory for AI agents that reasons about each user |
| Models | Neuromancer, Plastic Labs’ own reasoning models; works with any LLM |
| Plugins | Claude Code, Codex, OpenClaw, Hermes Agent, Pi and DeepSeek Harness, plus an MCP server |
| License | AGPL-3.0; self-host it or use Honcho Cloud |
| Pricing | $2 per million tokens stored; unlimited free reads; reasoning queries from $0.001 |
| Company | Plastic Labs, New York; founded 2023; $5.35M pre-seed |
| Best for | Developers building companions, coding agents and other agents that should know their users |
What Honcho does
Honcho’s homepage frames the problem simply: “Memory systems store facts. Users and agents are more than facts. And facts waste tokens. Delivering the context that matters requires reasoning.”
So Honcho does two things whenever an agent writes a message to it:
- It stores and indexes the message, so the history is searchable.
- It reasons and learns automatically. Plastic Labs’ Neuromancer models work out what follows from the conversation: “conclusions that follow, patterns across interactions, and hypotheses to test against new data.”
The agent then calls one method, context(), which “returns instantly curated reasoning plus conversation history.” Honcho’s pitch is that this gives agents the 10,000 tokens they need instead of the 100,000 they don’t, saving money and improving answers. One user quoted on its homepage said Honcho cut their contexts “down by 10-15% usually.”
Peers, not just users
Most memory systems assume one user talking to one assistant. Honcho models “peers,” which can be users, agents, game characters or groups, along with their relationships. Peers can be added to or removed from sessions, and each one’s perspective can be shared widely or kept narrow.
That opens up uses other memory systems handle awkwardly, which Honcho lists as group conversations, agents working with subagents, and characters in games whose memories evolve. Its site describes game characters that form opinions, track relationships and reshape the story as it goes.
Dreaming and on-demand reasoning
Honcho also reasons when nobody is waiting. Its “dreaming” runs in the background to identify patterns, test hypotheses, weigh conclusions, resolve conflicts and form a picture of each peer’s identity, without slowing the agent down. Standard dreaming comes with every workspace.
When an agent needs a specific answer, it can ask Honcho directly through a chat interface, choosing how much reasoning to spend:
| Level | What it does | Example question |
|---|---|---|
| Minimal | A single search | What’s the user’s name? |
| Low | Conclusions with surrounding context | When’s the user’s birthday? |
| Medium | Multiple searches and directed synthesis | Describe the user’s communication style. |
| High | Multi-pass analysis over time | How has the user’s sentiment evolved? |
| Max | Exhaustive full-history research | Build a churn risk assessment for the user. |
Plugins and integrations
Honcho has become especially popular as memory for coding and personal agents. It offers one-command plugins for Claude Code, Codex, OpenClaw, Hermes Agent, the Pi coding agent and DeepSeek Harness, a command-line tool, and an MCP server at mcp.honcho.dev for other AI tools. Its Claude Code plugin is pitched as carrying “how you work across repos and machines,” going beyond Claude Code’s built-in memory files.
Developers in the Hermes community are among its loudest fans. A post in the Nous Research Discord, quoted on Honcho’s homepage, reads: “I’ve tried them all, honestly. Mem0, QMD, Mempalace, and now Honcho. I freakin’ love Honcho, but I hate saying it.” These testimonials were chosen by Honcho.
Benchmarks
Honcho says it’s state of the art on three long-term memory benchmarks, LongMem, LoCoMo and BEAM, and that it’s “Pareto dominant” on accuracy, cost, speed and tokens. It publishes its evaluation code on GitHub so others can check the results. As with any vendor-run benchmark, the setup matters, and Honcho’s comparison pages against Mem0, Supermemory, Letta and Hindsight are its own.
Pricing
Honcho’s pricing, listed on its homepage, has two parts:
| Part | Price |
|---|---|
| Memory ingestion | $2.00 per million tokens stored, including Neuromancer reasoning |
| Reading memory back | Free and unlimited, about 200 milliseconds |
| Background reasoning | Included |
| Reasoning queries | $0.001 to $0.50 per query, by reasoning level |
New accounts get $5 of free credit. Startups that have raised less than $5 million can apply for $1,000 in credits and a year of subsidized pricing, and enterprise plans include forward-deployed engineers.
Open source and self-hosting
Honcho’s code is on GitHub under the AGPL-3.0 license, with about 8,100 stars. You can self-host it or use the managed Honcho Cloud. The AGPL is a strong copyleft license: if you modify Honcho and offer it to others over a network, you have to share your changes. That’s worth checking with your legal team before building a commercial product on a modified self-hosted version.
The company
Plastic Labs is a New York AI lab founded in 2023 by Vince Trost, its CEO, Courtland Leer, its COO, and Vineeth Voruganti, its CTO. According to its investor Variant, the founders ran into the problem while building Bloom, a tutoring chatbot: students weren’t learning well because the app didn’t understand who they were or how they learned. Honcho grew out of that.
In April 2025, Plastic Labs raised a $5.35 million pre-seed round led by Variant, White Star Capital and Betaworks, AlleyWatch reported, bringing its reported equity funding to about $6 million. The company describes itself as “an engineering-driven AI lab building at the intersection of machine learning and cognitive science.”
How Honcho compares
Honcho competes with other memory layers, including Supermemory, Mem0, Zep, Letta and Hindsight. Its distinguishing traits are reasoning at write time and in the background, the peer model for groups and multi-agent setups, a deep integration with coding and personal agents, and a fully open-source server. Supermemory leans more toward ingesting many content types and documents. Our guide to connecting Mem0 to an AI agent covers another rival.
Honcho is a building block for developers. If you want an assistant that already remembers you without wiring anything up, Carly has memory built in: it saves named memory documents that it reads and searches in later conversations, searches a knowledge base of files you upload by meaning, and keeps your to-dos, data tables and CRM in one workspace. That memory comes with you whether you use Carly inside ChatGPT, inside Claude or by text, and Carly also labels and sorts email across all your accounts, handles advanced scheduling, and makes and answers calls and texts.
Who Honcho suits
Honcho is a good fit if:
- You’re building a companion, tutor, coding agent or other agent that should understand its users over time.
- You use Claude Code, Codex, OpenClaw or Hermes Agent and want memory across sessions and machines.
- You need memory for group conversations, multiple agents or game characters.
- You want an open-source server you can run yourself.
Look elsewhere if:
- You need to ingest lots of documents and media rather than conversations.
- AGPL licensing doesn’t fit your plans for a modified commercial product.
- You want a finished assistant rather than a memory layer to build with.
FAQ
What is Honcho?
Honcho is an open-source memory system for AI agents from Plastic Labs. It stores conversations and uses its own reasoning models to draw conclusions about each user, then returns curated context to the agent with a single call.
Is Honcho open source?
Yes. Honcho is licensed under AGPL-3.0 and can be self-hosted, or you can use the managed Honcho Cloud.
How much does Honcho cost?
Ingestion costs $2 per million tokens stored, reading memory back is free and unlimited, and reasoning queries cost $0.001 to $0.50 each depending on depth. New accounts get $5 of free credit.
Does Honcho work with Claude Code?
Yes. Honcho has a Claude Code plugin that carries memory across projects and machines, plus plugins for Codex, OpenClaw and Hermes Agent and an MCP server.
Who makes Honcho?
Plastic Labs, a New York AI lab founded in 2023 by Vince Trost, Courtland Leer and Vineeth Voruganti. It raised a $5.35 million pre-seed round in April 2025.
Ready to automate your busywork?
Carly schedules, researches, and briefs you—so you can focus on what matters.
See what people say
"Before Carly, I relied on a Calendly link, but the whole process felt impersonal and not very professional. Carly changed that by handling all the back-and-forth, so I'm no longer stuck in endless email threads trying to line up schedules.
Now Carly reaches out to candidates, shares my real-time availability, lets them pick a slot, then sends a Zoom link and drops it straight into my calendar. She sends reminders to both of us before each call, which has significantly reduced no-shows and last-minute confusion.
On top of scheduling, Carly acts like a full executive assistant, sending me my schedule the night before so I can prepare for each call. It reminds me of the old x.ai assistant, but Carly is noticeably smarter, faster, and better suited to my healthcare recruitment business."
