Your company's brain.
Persistent. Searchable. Agent-readable.
Tribal knowledge dies in Slack threads. Decisions get re-litigated every quarter. New hires ramp in months, not weeks. Cerebro is the structured memory layer underneath your existing tools — Slack, GitHub, Linear, Notion, meetings, any webhook. Every signal becomes a typed, durable memory with provenance, links, and supersede. Every AI agent on your team reads from the same brain.
How the company brain remembers
A 60-second tour pulled from a real round-trip in our own dogfood vault. One decision in Slack on Monday, one PR on Tuesday, one agent query on Thursday — and the brain remembers all of it.
Slack message in #eng-platform: "Let's go with Postgres pgvector for now, revisit if recall drops below 80%."
→ A `decision` memory lands in /pending, awaiting one-click approval.
GitHub PR #482 merged: "Migrate embedding pipeline to pgvector".
→ Auto-linked to Monday's decision. Provenance preserved.
Your agent asks "why pgvector?" via the MCP server.
→ Returns Monday's decision + Tuesday's PR — full context, no re-asking.
Feed the brain from every system your team already uses
Connectors are Cloud-only — they're what turns vault-mem from a personal vault into a company brain. Self-hosters get the same MCP + keeper + markdown vault, minus the prebuilt ingest.
Personal brain or company brain?
Same memory format, same MCP surface, same keeper hygiene. The split is who reads from the brain — just you, or your whole team.
vault-mem · your personal brain
“My agent should remember what I decided last Tuesday, across all my projects.”
- Local markdown vault on your laptop — fully yours
- One developer, one machine, every AI tool you use
- MCP server (TS) + keeper (Py) + LanceDB index
- 9 memory tools · supersede · contradiction detection
- Telegram approval gate for destructive ops
- Skills export to Claude / Cursor / Windsurf / generic
- Zero hard cloud dependency · MIT licensed
Cerebro · your company's brain
“Our new hire shouldn't have to ask why we chose Postgres in 2024 — it's already in the brain.”
- Every system you already use, feeding one durable memory
- Connectors: Slack, GitHub, Linear, Notion, meetings, any webhook
- Multi-tenant Postgres + RLS · roles · SSO-ready
- Web UI: review queue, search, supersede, audit log
- Hybrid FTS + semantic search (pgvector + Voyage)
- Per-subject DPDP / GDPR erasure cascade built-in
- Every agent on the team reads from the same brain
Where does your data go?
Honest answer: the classifier sends each ingested item to an LLM to decide if it's a worth-keeping memory. Here's the threat model — and the three things you can do about it.
Your text never trains anyone.
The classifier uses Anthropic Claude. Anthropic's commercial API terms commit to not training models on customer traffic. We don't train on it either — we don't have a model.
BYOK is the default we're shipping for.
Plug in your own Anthropic, OpenRouter, or AWS Bedrock key. Your billing, your logs, your region. The LLM call leaves Cloud and lands in your provider account — we never see it. Ships before the first design partner.
Or skip the cloud LLM entirely.
Self-host vault-mem with a local Ollama or vLLM endpoint. The company brain stays inside your network. Same MCP surface, same memory format, same keeper hygiene — just no managed connectors.
Full data-flow diagram, third-party commitments, and roadmap mitigations: read the security doc →
Erasure is a button, not a ticket.
When a person asks you to forget them, you have one tool that cascades through every memory, every embedding, every full-text index, and the audit log. No SQL spelunking. No "we'll get to it." DPDP + GDPR aligned from day zero, by design — not bolted on.