feat: Honcho memory integration (opt-in)

[erosika]

Cross-session user modeling via Honcho. Hermes builds a persistent representation of who it's talking to -- preferences, patterns, context -- and carries that understanding across conversations. A model of the user.

Three integration points: prefetch user context into the system prompt each turn, sync exchanges for ongoing modeling, and a dialectic reasoning tool (query_user_context) for the agent to actively query its understanding of the user mid-conversation.

honcho_integration/ reads ~/.honcho/config.json as single source of truth. All imports lazy, all calls non-fatal. Optional dependency, zero behavior change when disabled.

Composes with USER.md

Both systems run in tandem. USER.md stays in the system prompt as the agent's local, curated snapshot of the user. Honcho prefetch is additive -- synthesized cross-session context layered on top.

When the agent writes to memory(target="user", action="add"), observations go to both USER.md (local file) and Honcho (cross-session reasoning model). MEMORY.md (agent's own notes) is untouched in all cases.

When Honcho is disabled, USER.md works exactly as before. Nothing changes.

Prefetch size is configurable via contextTokens in the global config -- caps how much user context Honcho surfaces per turn. No value = uncapped.

[erosika]

[teknium1]

Can you link me to more info on Honcho I havent heard of it to know what it adds vs the USER.md user profile that the agent already builds on their user

[erosika]

@teknium1

So USER.md is creating conclusions similar to how Honcho does, yes. To be clear, my PR currently runs alongside it -- I'll make one more commit to integrate them cleanly (still opt-in) if you'reinterested.
But with Honcho it's more than selective conclusions / RAG over said conclusions. You have configurable mapping of sessions, peers, and workspaces under the hood. I have ~100,000 conclusions about me captured via Claude Code over the last 3 months for example-- with Honcho, I can make that context immediately available to Hermes, compose chat turns for different levels of reasoning (.chat()) and summarization (.context()), and ingest that into the agent's immediate or long term awareness.

Important to note-- the reasoning Honcho does is powered by Neuromancer which was specifically fine-tuned for this task. (link to evals here)

There's also dreaming -- the reasoning model ambiently makes conclusions and fills in the gaps between sessions. It's a better way to form identity, all accessible in the cloud (remap anything on the fly) or self-hosted with your own model infra.

I'm using Honcho as a bridge between all of my agentic interactions and that's the most valuable part for me. The peer-peer architecture also lets you control which peers can make conclusions about other peers -- super helpful for agentic orchestration going forward.

Happy to discuss more -- erosika on Discord / @3rosika on Twitter.

Announcing Honcho 3 | honcho.dev

[teknium1]

Okay, after reviewing I like it a lot. I'll add more documentation on it after merge - Thanks!