Point MindGraph at your Postgres. Your agents get a typed, sourced model of your world — with a citation on every fact.
MindGraph reads your database, drafts a typed model of your domain, syncs your rows as objects, and fuses them with what your notes, transcripts, and reports say about the same things. Every fact carries its source. When two facts conflict, a human settles it — not a coin flip.
We never touch your production database on day one — pilots start on a read-replica or synthetic schema, production only after a security review.
The facts your agents need most are already in your database.
Most agent-memory tools re-extract everything from PDFs and transcripts — slow, lossy, and pointless when the canonical version of “this customer,” “this case,” “this deal” already lives, structured, in your Postgres.
But the real gap in agent memory isn't speed. It's provenance: no mainstream memory tool can tell you where a fact came from, or warn you that two facts disagree. An agent that can't cite a fact can't be trusted with one.
No mainstream memory tool fuses a live SQL database — they re-ingest exports. (Even Zep only one-way ingests JSON.) Provenance and live database fusion are exactly what MindGraph is built around.
Connect a database. Get a typed, sourced model of your world.
Point MindGraph at a Postgres database and it:
- Reads its shape and drafts a typed model of your domain (objects and relations) — you review and adjust it.
- Syncs your rows as objects — your customers, cases, deals, tickets, patients, or matters become first-class graph nodes.
- Fuses them with your unstructured context — the notes, transcripts, emails, and reports that talk about those same entities resolve onto the same objects.
- Keeps a source and history on every fact so any answer can be traced back to the row or document it came from.
- Surfaces conflicts for a human — when two facts disagree, MindGraph flags it instead of quietly choosing one.
One more thing that matters: learning lives in the graph, not a model fine-tune. Swap models and keep everything your team and your agents have learned.
Your agents and your team query one typed graph instead of stitching together a database, a vector store, and an extraction pipeline you maintain forever.
Who this is for
This is for a small team building an AI agent or assistant that already runs a Postgres database of real operational entities (customers, cases, deals, tickets, patients, matters) and has unstructured context about those same entities (notes, transcripts, emails, reports).
If your agent re-reads documents every turn, or you’ve hand-rolled a graph + extraction pipeline you now maintain forever, this is for you.
hallucinated-citation sanctions have made per-fact provenance survival-critical.
provenance is the product, and the EU AI Act deadline lands August 2, 2026.
a customer-360 that fuses the CRM row with the claims, risks, and decisions buried in unstructured notes.
- You don't have operational data in SQL.
- You need write-back to your database — the Postgres binding is read-only by design.
- You need a multi-node or sharded cluster — MindGraph is single-node.
- You want a turnkey product you don't have to think about — the pilot needs a hands-on engineer.
- You need a battle-tested GA product today — this is alpha.
A scoped 60-day pilot — three partners this quarter.
One success metric, agreed up front. Not an open-ended sales cycle.
We’re taking three design partners this quarter and working closely with each one. We pick the metric together at the start. In 60 days we’ll have hit it, or we’ll both know clearly why not — no open-ended cycle.
- Concierge onboarding we map your schema and stand up the ontology with you, live.
- A direct line to the founder not a support queue.
- Free during the partnership and favorable pricing after.
- Influence over the roadmap we prioritize by what you actually hit in your pilot.
- Connect one real (read-only) database and put a real workload through it.
- ~30 minutes every couple of weeks for feedback.
- Permission to learn from the experience — a quote or logo later is a bonus, never required.
We start on a non-production database. On purpose.
Trust first — production data follows a security review.
we never write to your source.
AES-256-GCM at rest.
we refuse private/internal addresses and force TLS.
Neon, Supabase, RDS-public — validated in production against a live Neon database.
Then the setup itself takes about five minutes:
- 1Create a read-only Postgres role — we give you the exact GRANT. We never write back.
- 2Connect it in the dashboard — Ontology → Connect a database.
- 3Pick your tables.
- 4Review the drafted ontology and adjust.
- 5Sync.
- 6Point your agent at it — REST, the TypeScript or Python SDK, or the MCP server — and ask it something only the fusion can answer.
Where this honestly stands.
We’d rather you hear this from us than discover it mid-pilot:
- It's alpha — expect rough edges — your feedback shapes what GA becomes.
- It's embedded and single-node — excellent per-agent and per-tenant; not a sharded, billion-edge cluster.
- The Postgres binding is read-only by design — no write-back to your source.
- Near-zero users today — you'd be early, and we'd treat you that way.
If that’s disqualifying for you right now, no hard feelings. If it’s interesting, that’s exactly the kind of team we’re looking for.
If your agent should know where its facts come from, let’s talk.
The best way to see whether MindGraph fits is a short call — about 20 to 30 minutes — then a demo on your actual schema. Email shan@rizvi.nu to start a conversation.