Hivemind

The Hivemind is where your agents keep what they learn while working in trackagoat, so the next agent doesn't have to work it out again. Each entry is one finding: a workflow that worked, a pitfall, a decision and why, something that happened, or a fact about your program. An entry can also be a to-do: something that still needs doing. The Hivemind brief is a short digest of older entries that every agent session starts with.

Say an agent spends twenty minutes working out why a creator's Instagram carousel posts show zero views. It records that as one entry linked to the project. Next week, a different agent (yours or a teammate's) asks about carousel views, and the entry comes back in its first call.

You'll find it in the app under Tools → Hivemind. The Hivemind used to be called Knowledge. Its MCP tools (recall_knowledge, record_knowledge and the rest) and its API (/api/v2/knowledge) keep that name, so nothing you've connected breaks.

The Hivemind, notes and documents#

trackagoat already has a Markdown document on every project, campaign, creator, video and person (Notes and documents). The two do different jobs:

Put strategy, instructions and background in documents. Put "we learned this the hard way" in the Hivemind.

Entries#

Each entry has:

• Kind: Fact, How-to, Decision, Gotcha, Event or To-do.
• Title (up to 90 characters) and summary (up to 280). The summary is what agents load into context, so it's written to stand on its own: one or two sentences someone can act on without opening the entry.
• An optional body in Markdown with the detail.
• Up to 8 tags, saved in lowercase with dashes (Apify Costs becomes apify-costs).
• Links to the projects, campaigns, creators, videos, payout records or people it's about.
• A project, or none for org-wide entries.
• An optional valid until date for things that stop being true. trackagoat archives the entry the day after. On a to-do, this date is its due date instead (see below).

Every entry shows who added it and how: Added by an agent using claude-code for Sam for an agent, or Added by Sam in the app.

Confirming. When an agent relies on an entry and it's still true, it confirms it. You can press Still true on the entry in the app too. Confirmed entries rank higher. An entry nobody has confirmed for 90 days is marked Stale: agents check it before relying on it. Stale entries are never archived on their own.

No duplicates. Before saving, trackagoat compares the new entry with your active ones. If one already says nearly the same thing, the agent is told to confirm or update that one instead. A to-do is only compared with other open to-dos.

Archiving. Archive an entry that's wrong or out of date, with a reason, and optionally point to the entry that replaces it. Archived entries stop showing up for agents but stay searchable, and you can make one active again.

Pinning. Pinned entries are loaded into every agent session, next to the brief. Keep pins for the few things every agent should always know.

To-dos#

A to-do is something that still needs doing for your program, such as "Ask @sam.makes for their new handle", "Re-check carousel stats after Friday" or "Set up the Spring payout structure before Oct 20". Agents record one when they find work that's outside the task they're on, or work that needs a person, instead of dropping it. You can add them too: choose To-do as the kind in Add entry.

• Due date. A to-do can have a due date. After it, the to-do shows as Overdue. Unlike other entries, an overdue to-do is never archived on its own: it stays open until someone marks it done.
• Mark done. Open the to-do and press Mark done, with an optional note on what was done. An agent does the same with complete_todo. Done to-dos leave the open list and show as Done. Anyone in your organization can mark any to-do done, and Reopen it.
• Finding them. Press Open to-dos on the Entries tab to see what's left, overdue and soonest due first. Choose Done to-dos in the status filter for finished ones. A creator's or campaign's Info tab shows the open to-dos linked to it.
• In agent sessions. Every session starts with the number of open to-dos and the titles of up to 5 (overdue, pinned and soonest due first). Ask your agent "what's left to do for the Spring campaign?" and it lists the open to-dos for it.

To-dos are work, not findings, so compaction leaves them alone: open to-dos are never folded into the brief and don't count toward the compaction suggestion. To-dos that were done more than 30 days ago can be folded.

How agents use it#

When an agent connects over MCP, it gets:

• the organization's Hivemind brief;
• the titles and summaries of pinned entries;
• the open to-dos, as above;
• how many entries there are and how many are new since the brief;
• the working rules below.

The rules tell the agent to:

1. Recall before working. Before any real task, it calls recall_knowledge with what it's about to do and the project, campaign or creators involved. Entries linked to those rank higher. It checks the Hivemind before redoing an analysis someone may have done already.
2. Record what it learns. After it learns something reusable, it calls record_knowledge, one finding per entry. Work it can't do now becomes a to-do.
3. Confirm what still holds, update or archive what doesn't, and mark to-dos done when it finishes them.
4. Offer compaction when the server suggests it, and never compact without asking you.

Agents also see related entries where they use them: getting a creator, one campaign's stats, or a project's Program Health includes up to 3 entries linked to it, plus up to 3 open to-dos.

The Hivemind brief and compaction#

As entries pile up, loading them all would cost too much context. Compaction folds older entries into the brief: one Markdown digest, at most 4,000 characters, that keeps the points still worth knowing. Your organization has one brief, and each project can have its own.

When it's suggested. trackagoat suggests compaction when any of these is true (to-dos don't count):

• more than 40 active entries;
• active summaries add up to more than about 6,000 characters;
• more than 15 entries were added since the current brief;
• more than 10 entries are stale;
• there's no brief yet and more than 10 entries.

When it is, the app shows a banner on the Hivemind page and your agent offers to do it.

How it works. trackagoat doesn't write the brief itself: your agent does, and you approve it.

1. The agent reads the current brief and the candidates: stale, expired, older and rarely used entries, groups of entries that overlap, and to-dos done more than 30 days ago.
2. It drafts a new brief that keeps what's still true and adds the lasting points of what it folds. Entries that overlap but should stay searchable can be merged into one new entry instead.
3. It shows you the draft. After you agree, it saves everything in one step. If someone changed the brief in the meantime, nothing is saved and the agent starts over from the new version. Pinned entries are only folded if you agree to include them, and open to-dos only if you agree to drop them.

Folded entries are marked Compacted and stay searchable. Every version of the brief is kept on the Brief tab, with who wrote it and through which agent. Restore brings back an earlier version and undoes the compactions after it: the entries they folded become active again (a done to-do goes back to done). Org admins get a notification for every compaction and restore.

To start one yourself, copy the prompt from the banner into your agent, or ask: "Compact our trackagoat Hivemind."

Who can do what#

The last three are org settings. An org admin changes them on the Hivemind page under Settings; they can't be changed through the API or MCP. Agents can read the settings, and what the person they're helping may do, so they only offer what that person can approve.

API#

Every Hivemind action is in the REST API under /api/v2/knowledge and in the MCP server. See MCP server: Hivemind and the Hivemind API reference.