Accelerating Agency.

What is Hivemind?

Hivemind is a self-hosted, offline server that gives a fleet of AI coding agents one shared episodic memory — behind a recall gate that returns nothing rather than guess, with memories re-verified against the code they describe. It is free and open-source under the Apache-2.0 license.

The project exists to solve one specific problem: when you run more than one AI coding agent on a codebase, they keep re-discovering the same things, and they have no honest way to tell a hard-won fact from a plausible guess — or a current fact from a stale one. Hivemind gives them a shared memory with a conscience: work is captured once and reused, nothing is served as settled fact until it has earned it, and what the code no longer supports gets flagged, mechanically.

Our editorial and trust stance

Every product claim on this website is meant to describe what the shipped code actually does — no more. We do not publish benchmarks we have not run, testimonials we do not have, or features that do not exist.

The same discipline that governs Hivemind's memory governs this site:

• Honesty over recall. If we are not sure of a number, we do not print a number. Where this site describes behaviour — the four trust states (quarantined, provisional, established, retired), the eight tools, the serve floor, the offline build — it reflects how Hivemind is designed to work, and the source is the ground truth.
• One number, and it is sourced. The only benchmark figure this site cites — poisoned-memory serve rate falling from 0.92 to 0.0 — was measured on the project's own adversarial trust bench. There are no testimonials, logos, or star ratings here because we do not have them yet, and inventing them would be exactly the failure Hivemind exists to prevent.
• Present tense means shipped. Anything not in the current release — like the enterprise tiers, marked coming soon — is labeled as upcoming, never written as if it already exists.
• Claims trace to code. The mechanisms described in the guide and FAQ map to real modules in the repository — the trust lifecycle and the tool definitions. Where we describe an interface, we link the source; the operating guide and admin docs (below) go deeper than this site does.
• Corrections are welcome. If something here is wrong or out of date, open an issue. Documentation that misrepresents the software is a bug, and we treat it like one.

The principles behind the design

Four commitments shape every decision: Hivemind stays on your machine, it earns trust rather than assuming it, it checks its memory against your code, and it would rather stay quiet than mislead. The guide shows how each one is built — this page is about why they are non-negotiable.

They are a direct reaction to how agent memory usually fails: silent staleness, poisoning, unbounded bloat, and a store no one can trust. A project that exists to fix those failures can’t cut the same corners in its own design, so these four are held as constraints, not preferences.

License and provenance

Hivemind is released under the Apache License 2.0, © 2026 Hivemind. You may self-host, modify, and use it commercially under the standard Apache-2.0 terms.

The canonical source of truth is the repository. The license text, the code that implements the trust lifecycle and the eight tools, the operating guide, and the changelog all live there:

• github.com/Hivemind-OSS/Hivemind — source, license, and issues.
• Operating guide (llms-full.txt) — the fuller reference this site summarises.
• HIVE-ADMIN.md — for operators running Hivemind for a team.
• CHANGELOG.md — what changed, release by release.

How to reach the maintainers

Email founder@hivemnd.tech — that reaches the project directly.

Because Hivemind is self-hosted and offline, there is no telemetry sending us data and no support portal — email is the front door. If you are evaluating Hivemind, the fastest path is to self-host it in a few commands and try it against your own fleet.