Shared LLM memory for teams
Everyone's agents, one graph, and a paper trail for whoever has to answer for what's in it.
Start a teamWorks with Cursor · Claude Code · ChatGPT · Gemini CLI · Windsurf · Antigravity and any MCP client · Full client list →
Every tool your team uses keeps its own context, and the one fix that would work makes you responsible for whatever ends up in the shared graph.
- Every tool keeps its own context; nothing syncs across people or tools
- Slack, Notion, and pasted rules are how context gets shared today
- The same facts get re-explained, slightly differently, every week
- Sharing what your agent learned means committing a rules file and hoping people pull
Today
Dev's architecture note: stuck in a Cursor session
PM's positioning doc: in a Claude chat that ended
Lead's sprint summary: in ChatGPT, not shared
Nobody's agents agree on the facts.
With Maindex Teams
→ Every agent, every person, reading the same facts
→ Decisions carry provenance: who, when, what it replaced
→ A new teammate's agent is current on day one
→ Screened on the way in, so shared memory isn't a shared attack surface
# merged recall: personal + team, labelled memory_search(q="token refresh") → mem-2c personal → acme:mem-4f acme · editor · approved
Same MCP connection your agents already use. Join a team and reads and writes scope to team collections automatically.
Guardrails
Shared memory is shared execution: a note your teammate's agent wrote is a note yours will act on. Secrets and API keys, injected instructions, and profanity are screened before publish. Screening cannot be disabled by an MCP client. Managers see the exceptions, not the traffic.
Accountability
Every write records who, which agent, when, and what it replaced. Corrections are new revisions. Managers can forget a memory, purge its history, or wipe a collection. You can destroy the content; the audit log records that you did.
Maindex Teams = shared graph + provenance + it survives contact with agents
- Personal: each person's own scope; managers cannot see it
- Team: shared collections everyone's agents can read
- Screen: every team write reviewed before the team sees it
- Prove: append-only log of every team action
Automated screening
Every team write runs server-side checks and agentic review before publish. Screening cannot be disabled by an MCP client.
Manager review queue
Approve, reject, or redact flagged writes. Redaction keeps a shadow copy only managers can see.
Roles per collection
Owner, manager, editor, viewer. Enforced in route handlers, service layer, and Postgres RLS.
Append-only audit log
20+ event types: invites, role changes, review verdicts, collection changes. Queryable by actor and date.
Managers see team activity and per-member daily read/write totals. Never personal memories. Security posture →
Three steps. No new tooling.
- 01 Create a team or accept an invite. Owners invite by email; joiners accept on their own account and plan.
- 02 Team collections appear in every MCP client you already use, alongside personal memory.
- 03 Managers watch the queue, not every write. Clean content publishes immediately; flagged writes wait for a decision.
memory_keep( headline = "We chose Postgres, see ADR-014", team = "acme", collection = "acme:decisions" ) → acme:mem-4f (review_status: approved)
Nothing to sign.
No contract, no annual commit, no minimum team size.
Who pays changes. What it costs doesn't.
- Everyone pays for themselves. No coordination, no approval, nobody waiting on anyone.
- One person pays for everyone. One invoice, one renewal, one thing to change when the team changes.
Join teams on any paid plan; create teams on Galaxy Brain ($49.99/mo). Full MCP + API access on every plan. Compare plans → · Teams FAQ →
Get shared LLM memory for your team
14-day trial. Create a team in about two minutes.
Start a team Card required. Cancel anytime.