Docs / Get started
How BrigadeStack works
What BrigadeStack is, its two parts, where its files live, and what runs in the background.
BrigadeStack keeps the agent tools you already use in step with each other. It is one command line tool, brigadestack, that writes plain files in your repo.
Two parts
Shared memory. An agent that learns something leaves a short note called a handoff. brigadestack ingest turns safe handoffs into memory cards, and any other agent tool in the repo can search them. With the backend running, the same cards reach other machines and cloud agents. See How shared memory works.
Checked work. This part is optional. Tasks, claims and checks record which commands ran for a task and how they exited, so you can see which agents, models and skills to trust. See Tasks, claims and checks.
Claude Code Codex Cursor OpenCode ... agent tools
\ | | /
v v v v
brigadestack (the CLI)
/ | \
v v v
memory/cards/ .brigade/ built-in engines
handoff inboxes tasks, claims, code graph,
per tool check records evidence ledger
Where files live
.claude/memory-handoffs/,.codex/memory-handoffs/and similar: one inbox per agent tool, where handoffs are written.memory/cards/: the memory cards thatingestpromotes..brigade/: tasks, claims and check records for this repo. It is local state and stays out of git.AGENTS.md,CLAUDE.md,SAFETY_RULES.md: instructions the agent tools read at the start of a session.
The code graph and evidence engines are separate programs. The CLI starts them as child processes and keeps their output and exit status.
What runs in the background
care installwrites scheduled jobs for the repo, such as ingest every 30 minutes and a daily memory check.- The backend and the web views run as serve commands.
- In 0.31.0 no scheduled job publishes cards to the backend.
brigadestack memory hub pushdoes that. - On Windows,
care installprints the Task Scheduler commands and you create the tasks yourself.
0.32.0 beta: brigadestack care install --entry lesson-review adds an optional job that runs every five minutes. It reviews a work session that went idle after enough tool calls and looks for a lesson. Automatic admission of those lessons is held until the backend anchor lands, so it does not change what your agents read yet.