Integration
buzz agents, with a memory.
buzz puts humans and AI agents in the same rooms. What it does not give those agents is a memory: every prompt starts cold, and what an agent worked out on Tuesday is gone by Wednesday. Captain Memo fixes that, and it needs no work on either side — the agents buzz already runs speak MCP, and Captain Memo is an MCP server.
The mechanism
Nothing here is buzz-specific.
That is the point, and it is worth being precise about. There is no buzz plugin, no
connect buzz, and no integration between the two projects. The join happens one
level down: buzz's harness runs an ACP agent, that agent is also an MCP client, and Captain
Memo is the MCP server it talks to. Wire the agent and the memory comes with it.
buzz runs an agent
Its buzz-acp harness drives goose, Codex or Claude Code over ACP on stdio.
That agent loads MCP servers
All three are MCP clients in their own right. Captain Memo registers as one of their servers.
The memory is already shared
One local corpus, on your machine. What the agent learns in a buzz channel, your terminal recalls later.
Setup
Install once, then wire the agent.
Install Captain Memo (it needs Bun 1.1.14 or later), then wire whichever agent your buzz relay drives.
git clone https://github.com/kalinbogatzevski/captain-memocd captain-memobun install./bin/captain-memo installcaptain-memo connect goosecaptain-memo connect codex
Claude Code needs no connect — the plugin install already wires its MCP server,
hooks and skill. connect is idempotent: run it twice and the second run reports
already registered and leaves the config byte-identical.
What the agent gains
It stops asking you the same thing.
Search over everything the corpus holds, the ability to write findings back, and full recall of a past session's detail. Memory written by your terminal is readable in a buzz room, and the reverse, because it is one store and not two.
The memory is local to the machine the agent runs on. Wire the agent on your laptop and that laptop's corpus is what it reads. buzz can also deploy agents onto remote substrates, and an agent running in a pod somewhere else has no Captain Memo behind it and no memory of yours. Nothing here changes that.
Only three of buzz's surfaces are covered, because those are the three its harness drives. And Captain Memo does not read buzz's event log, post to channels, or know that buzz exists — it is the agent's memory, not the workspace's.
buzz is young and moving fast. This page deliberately documents the MCP side, which is stable, rather than buzz's own surface, which is not.
More than one machine
A laptop, a desktop, a server.
Everything above is free and stays on one machine forever. If you end up running agents on several and want each one to be able to ask the others what they know, that is Captain Fleet, a separate commercial relay. It is not a trial and nothing on this page is held back for it. Memories are shared, never merged: each captain keeps its own memory of its own project, machine and time, and nobody copies anyone's whole memory into one pooled truth. If SAMI noted at 10:05 that the apples are 11, your captain answers “As per SAMI, the apples are 11”, not “the apples are 11”. A newer note on another machine never overrides your own, and SAMI's apples may be different apples.