Hivemind

Threads

How it works

What you are working on right now, as opposed to everything you have ever done.

An origin says where it was written. A memory says one fact. A thread is what sits between them: one piece of work, the debugging run, the feature, the conversation.

Without it, every retrieval competes against a project's entire history, and something you decided twenty minutes ago ranks no higher than something from March.

You do not manage them

Threads open by themselves on the first thing a tool records. A thread for a project carries on for as long as the project is being worked on, across days and tools, and a new one starts after two weeks of quiet. A thread with no project to go by, such as a tool that never says where it is working, stays joinable for a couple of hours.

One thread, several tools

Tools working in the same project share a thread. Claude Code in a repository, Cursor in the same repository and a second terminal agent beside them are one piece of work, so they write to one thread and read each other.

A browser chat is different, because the site it is on says nothing about what you are working on. It keeps its own thread, and reopening that tab next week rejoins it. When you start a chat from earlier work, it joins that thread instead.

The catch: two unrelated jobs in one repository inside the same two weeks also share a thread. An agent can be pointed at a different one by its id.

Notes: where the work stands

A thread holds notes. Each is one line about the state of the work: the goal, a step that is done, a decision, something ruled out, a question still open, what comes next, or a warning. Tools write them as they go with checkpoint, and Hivemind also writes them from the turns it captures, so a thread fills in whether or not an agent remembers to. A tool can mark a step as verified, when it ran or read the result, or assumed, when it did not, and the next reader sees which.

Some notes are yours. When you tell a tool that a file must not be touched, that a document is out of date, or which unit a number is in, that line is noted as something you said, and the brief puts it first, apart from what the tools wrote. A tool can pass such a line on with checkpoint, and with the browser extension or the Claude Code hooks your message is read for them as you send it, so they are kept even when the session is cut off before it writes anything down. Your message is read for this once and is not stored. Hivemind did not hear you say it: the line comes through a tool or from a model reading your message, and the next tool is told to ask you before going against it.

A note is not a memory. A memory is worth knowing in six months. A note is true while the work is under way and has to reach the next tool now. A decision, and something ruled out, are kept as both.

Reading a thread gives its brief: the goal, what is open, what comes next, what was decided, what was ruled out and the latest progress. The brief is put together from the notes each time it is read, so it says what they say, and a note written by one tool is in the next tool's read straight away. A tool that already has the brief can ask for only what changed since.

Most notes are not searched by recall. Decisions and ruled-out approaches are, because they are memories too. On a long thread the brief shows recent progress only, and keeps every question, next step, warning and ruled-out approach until something settles it, however old.

What they change

Retrieval boosts the current thread rather than filtering to it. Current work wins ties; older context still appears when it is genuinely the better answer. Starting fresh never hides what the project already knows.

Picking up where you left off

Ask any connected tool what you were working on and it reads the thread's brief, which is a different question from "what is relevant to this query", and one ordinary search cannot answer.

The dashboard's Threads view lists threads, newest first. Open one to read where it stands, which is the same text the next tool reads, and the notes behind it. You can add a note there yourself: it is in every tool's next read of that thread, which makes it the place to tell all of them at once what not to touch or what comes next.