# Honcho Honcho is an open-source memory and personalization library for AI agents, with an optional managed service. It is what [[Hermes Agent]] uses for dialectic user modeling, and the closest thing the open-source agent ecosystem has to a *reasoning-enhanced* memory layer (as opposed to plain RAG). The framing is structural; agents need to track entities (users, peers, ideas, agents themselves) over time, not just retrieve text. Honcho models entities as first-class objects whose representations evolve across sessions. ## The four primitives - **Workspaces**; application-level containers that scope everything else. - **Peers**; persistent entities that change over time. Users, other agents, projects, anything an agent maintains a model of. - **Sessions**; bounded interaction threads. Time-boxed and topic-boxed. - **Messages**; the data units that trigger the reasoning step. When messages arrive, custom reasoning models generate **representations**; structured conclusions about peers stored for later querying. The agent reads these representations as context, not the raw messages. ## Why it is different from RAG Plain RAG retrieves what was *explicitly stated*. Honcho extracts what is *latently true* about a peer through reasoning over the message history. Concretely, RAG can retrieve "user said they prefer concise answers." Honcho can derive "user gets frustrated with long preambles" without that ever being stated, by reasoning over patterns across sessions. The library calls this "rigorous, compute-intensive thinking that humans struggle with, instantly and consistently." That makes it valuable for any agent whose job depends on understanding the user well over time; personal assistants, coaches, long-running professional agents. ## How it ties into the wider stack - [[Hermes Agent]] uses Honcho for the user-modeling component of its persistent memory. - The pattern generalizes; any agent with a long-running relationship with a user benefits from this kind of structured memory layer. - Distinct from short-term context management (which is the model's context window) and from skill memory (procedural memory of how to do things). ## License and access Open-source library; managed service available at app.honcho.dev. Operated by the Honcho team, with first-class integration support for [[Claude Code]] and the broader agent ecosystem. ## References - Documentation: https://docs.honcho.dev/ - App: https://app.honcho.dev/ ## Related - [[Hermes Agent]] - [[AI Agent Memory]] - [[Personal Context Management (PCM)]] - [[Retrieval-Augmented Generation (RAG)]] - [[AI Agents]]