Astrocyte
Memory for every agent.
The memory layer. Section-grain recall, multi-bank isolation, full retain / recall / reflect / forget / compile API. Bench-validated against LoCoMo and LongMemEval.
Open source
Position
Collective Intelligence
What happens when humans and AI agents think, decide, and act together — and remember what they decided.
Most "multi-agent" systems today are multi-call LLMs with no shared state.
Each turn forgets the last decision. Each agent has no idea what its peer just said three messages ago. Each session starts cold.
That isn't collective intelligence. That's parallel autocomplete.
Real collective intelligence — what humans have done for millennia in teams, councils, and committees — requires memory. Decisions accumulate. Context carries. Tomorrow's discussion starts where today's ended.
Memory is the substrate.
LLMs are stateless. Context windows compress under load. Vector search retrieves snippets, not structured decisions. None of that is enough to sustain a team's thinking across weeks or a company's thinking across years.
Collective intelligence needs more than retrieval:
- Memory of decisions, not transcripts. What a council concluded matters more than what was said.
- Governed recall. Every read is access-controlled; every write is audit-logged.
- Multi-bank isolation. One team's collective context never leaks into another's.
- Structure-preserving retrieval. Recall across many sessions without flattening discourse into bag-of-snippets.
Build this once. Use it everywhere humans and agents deliberate together.
Three principles
1. Humans-in-the-loop.
Agents propose; humans approve. Collective intelligence isn't autonomous AI — it's humans and agents with shared context. Humans set direction; agents do the work; humans review and decide.
2. Agent-as-peer.
In a council, an AI agent is a participant, not a tool. It speaks, votes, dissents, gets overruled. The protocol is the same whether the speaker is human or model. Treating agents as peers is what unlocks collective intelligence; treating them as tools reduces them to fancy autocomplete.
3. Governed recall.
Memory without governance is a compliance breach. Every recall is access-controlled, every retain is audit-logged, every bank is isolated. Collective intelligence at scale needs the same guardrails as any other system that touches user data.
How we deliver it
One open-source stack. Three ways to run it.
Synapse
Self-host your Collective Intelligence.
The open-source deliberation layer. Multi-surface group chat — web, desktop, mobile, MCP — where humans and AI agents run councils on top of Astrocyte memory.
Open source · Apache-2.0 · BYOK
Cerebro
Collective Intelligence for your team.
The managed cloud version. Business tier for SMEs (hosted, OIDC SSO, audit log). Enterprise tier for large and regulated organizations (SAML, SCIM, DSAR, SOC 2 / GDPR / PDPA / HIPAA).
Paid · Managed by Odeon
Bring Collective Intelligence to your team.
Three ways to run it: open source DIY, hosted by us for teams, or fully managed for enterprises.