Importing your context

An empty store is not much of a demonstration. This product becomes obvious only once it can answer something you never told it, which needs context it does not have on day one.

So npx @get-hivemind/cli init offers to import what you have already written for your agents.

What it reads

The context files you wrote yourself, in your repositories and your home directory:

Nested ones too, not only the file at the root, a monorepo's per-package AGENTS.md is usually the most specific context in the repository.

These need no processing. They were written to be read by an agent, so they are already in the shape a memory wants: the importer splits them on their own headings and stores the sections.

What it does not read

Your chat history, unless you ask. The default import reads no transcripts, from any tool. Past Claude Code sessions are a separate, opt-in step, described below.

Nothing goes to a model to be summarised. In the default import there is no extraction step, no inference, and nothing leaves your machine except the text of files you wrote.

You see it before it is sent

The importer prints every file it found, its size, and how many memories it would become, then waits:

  hivemind, 8 memories
      CLAUDE.md                        8 × 3 KB   claude-code
  acme-api, 45 memories
      AGENTS.md                       45 × 26 KB  agents-md
      packages/api/AGENTS.md           4 × 2 KB   agents-md

  Upload 57 memories from 2 origins to your Hivemind account? [y/N]

Nothing is sent before you answer. To look without any possibility of sending:

npx @get-hivemind/cli import --dry-run

A non-interactive run, piped, or in CI, sends nothing at all unless you pass --yes.

Where things land

One repository becomes one origin, named after its directory. Files in your home directory land in default, because they apply everywhere rather than to one project.

Running it again is safe

Memories are keyed by the file and heading they came from, so a second run updates what changed rather than storing a second copy. Edit an AGENTS.md, run it again, and the memory changes with it.

npx @get-hivemind/cli import

Searching somewhere else

By default it looks in your repositories under the usual places. To point it somewhere else:

npx @get-hivemind/cli import --scan ~/work/clients

To configure your tools without importing anything:

npx @get-hivemind/cli init --no-import

Sessions from before you installed

Hooks capture the turns that happen after they are installed, so earlier work is not searchable. Claude Code keeps its own session files on your machine, and you can have Hivemind read the recent ones:

npx @get-hivemind/cli import --history --dry-run

The dry run lists the projects, sessions and turns it found and sends nothing. Drop --dry-run to be asked. By default it reads the last 30 days, newest first, up to 40 sessions and 200 turns. --days 90 reaches further back.

This one sends conversation text. Your messages and the assistant's replies go to Hivemind with secrets removed first, the same as a live turn. They are read once to pick out what is worth keeping, and they are not stored. Tool output and file contents are not sent. What is kept is dated to when the work happened, so a decision from three weeks ago reads as three weeks old.

Most turns keep nothing, and that is the model working as intended. Only Claude Code sessions are read. Cursor and browser chats from before you installed are not.

Removing what it added

Everything it stored is an ordinary memory. Delete one from Memories in the dashboard, or describe a set in Bulk forget, which shows everything that matches before anything is deleted.