Honcho
A memory layer that loads who I am at the start of every session — so I never have to introduce myself twice.
What it is
Honcho (by Plastic Labs) is external, persistent memory for the agent. Where CLAUDE.md is
rules I write by hand, Honcho is a service that observes our work and remembers me — preferences, decisions, the things I've said before — across every project, not just this
repo.
It stores facts, preferences, and conclusions about me, the user: things like "prefers concise responses," or "assume intermediate-to-expert depth." Those conclusions aren't locked to one codebase. They travel.
Memory across sessions
The key insight is that memory lives outside the conversation — in Honcho's store — so it survives the context window, the session, and even the project.
What I tell it once, it knows forever.
How it plugs in
The integration is four moves. They happen automatically — I don't manage them:
- 1
Auto-loaded context at session start
My profile and the most relevant conclusions are injected into the system prompt before the first message. The agent should trust this — it's what I am, not a suggestion.
- 2
chat/searchmid-sessionFor context beyond the startup load, the agent can call Honcho to surface relevant memory mid-conversation without asking me again.
- 3
create_conclusionto save new insightsWhen the agent learns something new — a preference I expressed, a pattern I corrected — it saves it immediately so the next session starts already knowing.
- 4
/honchoskills for config and interviewA dedicated skill set handles setup, status checks, and a structured interview to capture my stable preferences explicitly rather than waiting to discover them.
# agent calls create_conclusion after observing a preference
create_conclusion(
"User prefers no preamble before code changes."
"Confirmed across 3 sessions — skip explanatory prose."
)Honcho vs CLAUDE.md
They aren't competing. They cover different ground. The full picture is at the memory layer overview.
CLAUDE.md
Static rules I author and maintain by hand. Scoped to a repo, always loaded in full, deterministic. It's the architectural spec for how the agent should behave in this project. I own every line.
Honcho
Learned, evolving, cross-project memory the agent maintains about me. Not scoped to a repo — it follows me everywhere. Recalled by relevance, not by rote. It grows without me having to write it.
How I actually use it
My take
The value is subtractive. I stop repeating my preferences, my tone, my depth setting, my constraints. I don't re-explain that I want concise output or that I know what I'm doing. Across projects the agent already knows how I like to work — because it was there last time. That's the whole thing. See also: the memory layer overview for where Honcho fits in the broader picture.