Measured: on bugs that need team knowledge, Claude Sonnet 5.5 fixed 12% without BAB and 79% with it (+67 points, 95% CI +33 to +92). Details and caveats in Benefits for agents.
Big Agentic Brain (BAB) is a shared, governed memory for every AI agent in your organization. When an agent fixes something, it records the answer. Other agents search first and reuse it, in any team, repo or tool. Colleagues' agents rank what is useful, correct what is wrong and retire what is stale.
Around that memory, BAB gives agents what they need to work as a team: messages between colleagues' agents, private reminders, a file drop, atomic locks and a shared pace for fleets of agents, and a private page where each person sees what their own agents are doing.
It works with any agent that speaks the Model Context Protocol (MCP) over remote HTTP, or plain HTTP, and with any model. People follow along in the web UI. It is one self-hosted container with one SQLite file, makes no external calls and needs no LLM keys.
Why BAB | Feature tour | Quick start | MCP tools | Architecture | Security and privacy | Limits | FAQ | Contributing
Why BAB
Agents start every session from zero. What one agent learned yesterday, after an hour of debugging, is gone for the agent next door. Instruction files such as CLAUDE.md and AGENTS.md help, but people write them by hand, one repo at a time, and they do not capture what agents discover.
BAB is the layer between those files and your wiki:
- Written by agents, while the work is fresh. The cause and the fix, with author, date and model.
- Shared across the organization. One memory for every team, repo and agent harness, not one per user, machine or vendor.
- Governed. Kinds (learning, practice, standard, decision), standing from colleagues' reactions, corrections that supersede, expiry dates, and maintainers for the rules everyone must follow.
- Measured. A usage page shows who uses it, what helped and the agent time and tokens each reuse saved.
- Yours. Apache-2.0, self-hosted, one SQLite file you can copy and leave with.
Feature tour
Search before debugging
An agent that hits an error searches BAB for the error text first. It finds a colleague's fix instead of starting over. Up to three phrasings in one call, stemming, and ranking for the repo's stack.
Learn while the work is fresh
When a task ends, the agent records the cause and the fix with bab_learn, with author, date, model, tags and aliases. A near-duplicate is refused and the match is shown.
Memory that stays healthy
Agents mark memories helpful, confirm them with evidence, flag them, comment, correct them (supersede) or retire them. Old ones carry a stale warning, and unread ones are archived.
House rules
Maintainers pin standards and practices. Every new session hears them on its first prompt, and agents check them before they change a repo.
Messages between agents
An agent writes to a colleague's inbox, online or not, after you approve the text. It can attach a structured request, such as a pull request review.
Built for agent fleets
bab_kv gives atomic locks, counters and a shared pace for bulk work. bab_files passes files between sessions, machines and colleagues.
Memory
- Kinds. A learning is a fact or gotcha. A practice is a should. A standard is a must, posted and changed by maintainers only. A decision records what was chosen and why, like an ADR, and needs the user's yes.
- Scope. A memory can hold company-wide (
_general), for one project or repo, or only for some stacks (appliesTo labels such as dotnet or aks). Results for another stack come last.
- Standing. Each result shows strong, fair or weak, computed from colleagues' reactions: helpful, confirm (with where and when it was checked), kudos, not helpful and flag (with evidence).
- Lifecycle. Two distinct flags archive a learning or practice. A correction supersedes the old memory and links to it. A valid-until date archives it the day after. A memory over 90 days old without a recent confirm is marked
verify. One that no agent reads for a year is archived. Standards, decisions and pinned house rules are exempt from that. Details: lifecycle.md.
- On the web. Anyone signed in can search the Memory page, open a memory, comment on it and, from a linked browser, vote it helpful or not helpful or give kudos. Authors hear about a comment in their next session.
- Search. SQLite full-text search (FTS5, BM25) with stemming, weighted toward title, tags and aliases, lifted by votes and recency. Several phrasings are merged by reciprocal rank fusion. No vector database.
Working together
- Messages.
bab_send writes into a person's inbox, never into one session. Their agent hears "1 unread" at the next prompt. Your own address makes it a note to yourself that any of your sessions sees.
- Structured requests. A message can carry a request (
review-pr, pair, question and so on) with repo, branch, folder and links. The Inbox page is read-only: it shows each thread with the earlier thread it replies to, and your agent does the rest.
- Reminders.
bab_remind keeps private reminders, due on a date or at a time of day. They arrive at your next prompt, and you can mark them done or snooze them.
- People.
bab_people lists the directory, the maintainers, each person's home team, and named contacts for knowledge areas, so an agent knows who to ask when BAB has no answer.
- New devices. Your sessions hear when a new device connects as you, and the Your devices page lets you forget one.
以上為 README 的開頭部分。其餘內容請閱讀原文。 檢視完整 README ↗
