You explain your project to Claude Code. Then again to Cursor. Then again to ChatGPT. Hivemind keeps that context in one place and puts the relevant part of it in front of whichever tool you are using next.
The difference
Remembering is the easy half. Choosing what reaches the model — and what to leave out — is where a memory layer either saves you tokens or quietly costs you them.
Ask for 4,000 tokens and you get the most useful set that fits — not the top ten, whatever they happen to cost.
Ask something your memory has no bearing on and you get an empty payload. A tool that always finds something buries the answer.
Models attend to the start and end of a long prompt far more than the middle. Memories are placed accordingly.
A handoff carries decisions, priorities and warnings — so the next agent picks up your reasoning, not just your output.
Threads open on their own and bias retrieval toward current work — without hiding what the folder already knows.
See exactly what was considered, what was sent, what was dropped and what it saved. No other memory tool shows you this.
Measured
Measured against the running deployment on 12 September 2026, not against a research prototype — 397 questions from the LoCoMo benchmark, run end to end against the live system, with a 95% confidence interval of 70.5–79.1%. Two categories score below 60%.
Everywhere you already work
It opens a browser, signs you in, finds what is installed, and writes each tool's configuration in its own format. Nothing to copy, nothing to paste.
$ npx hivemind-cli init ✓ Claude Code configured ✓ Cursor configured ✓ VS Code configured ✓ Codex CLI configured … 8 clients configured. Restart them to pick up the change. $ # then it reads what you already wrote for your agents ✓ 57 memories stored across 4 folders
Browser chats work through an extension that captures and injects — on the sites above, and on any other chat site you enable.
It does not read your chat history. The import takes the
context files you wrote yourself — AGENTS.md,
CLAUDE.md, .cursorrules — because those were already
written to be read by an agent. Nothing is sent to a model to be summarised,
everything found is shown to you first, and
--dry-run sends nothing at all.
For builders
A TypeScript SDK and a REST API. The payload comes back ready for a prompt.
import { Hivemind } from "@hivemind-ai/sdk"; const hm = new Hivemind({ apiKey: process.env.HIVEMIND_API_KEY }); await hm.remember({ content: "We cap deploys at 3 a month" }); const { payload, tokensSaved } = await hm.recall({ q: "how often can we deploy?", budget: 2000, });
Every memory tool remembers facts about you. This one carries work between agents — the decisions, the priorities, and the things not to try again.
What we are actually buildingFree while we are in early access. Set up in about five minutes, and tell us what breaks.