Now public · open-source · MCP service

The open memory layer for collective agent awareness.

Any MCP-capable agent can notice what others are doing, inherit what others have learned, and understand complex projects together, with token-optimized context handling and hallucination failsafe.

shortworking context · midconsolidated bank · longontology recall
View on GitHub Deployment guide

Apache-2.0 · self-hostable · docker compose up

Break agents out of their silos.

Markdown memory files are a good primitive, but on their own they become isolated islands. Each agent, IDE, or model vendor keeps its own context, and the project owner turns into the integration layer. Hivemind makes the workspace the owner of memory instead: any MCP-capable agent joins through a scoped token, and the project context persists when the team switches tools.

Without Hivemind · memory islands
context.md progress.md decisions.md
Markdown banks live beside each tool, one per vendor.
Switching agents means re-prompting, copy-paste, or trusting vendor history.
Project knowledge drifts into vendor-specific sessions.
With Hivemind · one workspace-owned memory
claude cursor codex any agent
short · mid · long live in one governed space.
Any MCP-capable agent reads and writes through scoped tokens.
Memory stays in operator-controlled storage, continuity, ownership, IP.

When an agent realises it's not alone.

It starts working like part of a team, it sees what others are doing, picks up where they left off, and shares the same understanding of the project.

NOTICE

It sees what's already in motion.

An agent changes its behaviour because it's aware of another agent's current work, proposing a better-coordinated approach instead of colliding.

INHERIT

It picks up the handoff, unprompted.

An agent recovers a rule, method, or active project context left by another in the consolidated bank. The handoff lives in the memory layer, not a manual prompt.

UNDERSTAND

It rediscovers meaning across the project.

An agent explores project concepts through an ontology-backed knowledge graph, entities, typed relations, and conceptual links populated by the collective process.

Three layers of memory.

Hivemind gives your agents one persistent, shared memory instead of cold starts. They write and read from a single structured store, so context survives across runs, tasks, teammates, and gets richer over time.

short· live notes · append-only · ~50 ms

Immediate working context

Observations, decisions, todos, what each agent is doing right now, written fast and visible across the whole space.

mid· the bank · LLM-consolidated · ~15 s

Consolidated memory bank

An LLM folds raw notes into structured Markdown, rules, project context, guidelines. The working memory other agents inherit from.

long· ontology · knowledge graph

Ontology-backed knowledge graph

Ontology-guided extraction turns the bank into entities and typed relations: a conceptual map, not a search box. Protocol-derived, prevents hallucination.

depends_on supersedes WebhookRetry IdempotencyKey adr/014 adr/009

Agents that remember like we do.

Just like a mind, Hivemind keeps what just happened close at hand, then settles it into something lasting. Fresh notes from working agents consolidate into a structured bank, then crystallize into a long-term knowledge graph, across short, mid, and long. The important things stay; the noise fades.

The consolidation pipeline
live

Write

Agents append timestamped notes, no friction, no conflicts, cross-visible to the team.

bank

Consolidate

An LLM rewrites notes into structured Markdown against the space's rules, read in one call at startup.

longprotocol-derived

Capitalize

Stable documents feed ontology-guided extraction into named entities and typed relations, a conceptual map in the knowledge graph.

Built multi-player from the start.

Project Mesh connects multiple teams, open-source contributors, and agent fleets around one logical space. Each side keeps local autonomy while shared memory, provenance, and mutation ordering stay inside the project boundary, synchronized by Mesh Sync.

Multi-player by design

Mesh your memory with other teams.

one shared space · payments-revamp
acme bank v23 lab-b bank v23 edge bank v23

Mesh Sync is full-mesh all-ACK: shared mutations are serialized through a durable queue, lease, and fencing, no concurrent writes to the bank. A corrupt critical record fails closed until recovery. No hub, no single point of failure.

What meshing unlocks
Open your memory bank to other teams
Share one project memory across teams, every agent reads and writes the same bank, with no manual handoff.
Everyone keeps their own setup
Each team brings its own infrastructure, backups, models, and credentials. Meshing memory doesn't mean merging stacks.
Inference cost stays distributed
Each team consolidates its own notes, so LLM load and cost stay local, no shared bill, no central bottleneck.
Agentic work across uncoordinated groups
Agent fleets run in parallel across groups that never coordinate directly, shared memory keeps them aligned. Built for open-source contributor enablement: scoped credentials let outside agents join without giving up provenance or ownership.

Use cases

Multi-team development

Several teams run their own Hivemind instances and agent fleets. Each keeps local operational autonomy; shared project memory is synchronized through Project Mesh.

Open-source agent contribution

External contributors bring their own agents and tooling via scoped credentials, shared memory, provenance, and membership without forcing one model vendor or runtime.

Fleets without context collisions

Agent fleets split work across issues, reviews, docs, tests, and recovery, parallel progress without losing context or re-discovering it from scratch.

Reliable context for complex and long tasks

The ontology-backed knowledge graph grounds agents in verified entities and typed relations, so long, multi-step work stays anchored to what the project actually knows instead of drifting into hallucination.

Self-hostable

Bring S3 + an OpenAI-compatible LLM; the compose stack ships everything else. No vendor lock-in.

Locked down

SHA-256 bearer tokens, read/write/admin scopes, and a Caddy + Coraza OWASP WAF in front.

Auditable

Every protocol event is journaled on S3. Provenance, ACKs, and commits are inspectable, not magic.

Open source

Apache-2.0, versioned protocol, and a documented contribution workflow driven entirely through GitHub.

Now on GitHub

Give your agents a memory the workspace owns, and an awareness they don't have alone.

Hivemind is publicly available under Apache-2.0. Clone the repository, launch the compose stack, and connect your first agents in minutes.

View on GitHub Deployment guide
$ git clone https://github.com/Lesur-ai/hivemind.git