We scored our own memory system. Here's the receipt.

Every belief has a paper trail

When a fact enters CLS++, a fact changes, a contradiction is detected, or a memory is served into another model, it writes one immutable, hash-chained event. The chain is per-tenant and verifiable offline with a script we publish, so we cannot quietly rewrite history and re-hash it. An auditor does not have to trust our word. They can check the math.

Six surfaces read from that one spine. An audit ledger with before and after on every change. Lineage graphs that show how a single fact evolved, Pune to Bangalore to a corrected address, as a sequence you can walk. Belief timelines with the contradiction that retired each stale value. Red, yellow, green privacy where a red fact, a salary or a credential, is blocked from ever crossing into another LLM at the API layer, not in the interface where it is too late. An admin resolver that answers the question a breach response actually asks: given this leaked key between these two timestamps, exactly which users and which facts were touched. And full history controls where deletion means cryptographic erasure that keeps the chain verifiable, not a silent row drop.

So we scored ourselves, out loud

Claiming to be compliant is easy. Proving it is the job. So we built a control catalog, 29 controls mapping the clauses of GDPR, India's DPDP Act 2023, SOC 2 and the OWASP LLM Top 10 to the exact surface that satisfies each one, and we score it live against the running deployment. Not a slide. A page inside the product, refreshed on real state, that any admin can open and any auditor can be walked through line by line.

Here is where that scorecard stands today.

Two honest notes on that number, because the honesty is the point. It is a self-assessment against our own control catalog, not an external certification, and we say so on the page. And the score is scored on the deployment, not on a wish. Wire an off-host immutable anchor and a SIEM stream and it climbs. Leave them off and it does not. The scorecard reflects what is actually true today, which is the only kind of scorecard worth publishing.

The 13.8% we have not closed

This is the section a marketing page never has. Here is exactly what is amber and why, in our own words.

• partial

  GDPR Art. 17, erasure. Memory items are cryptographically erased today. Nulling the same content out of the raw conversation archive is next.

• partial

  SOC 2 CC6, access management. Read-only auditor role and admin gating shipped. Enterprise SSO and SCIM are on the roadmap, not in the box yet.

• partial

  External anchoring. The hash chain anchors to disk today. Off-host WORM or S3 object-lock, which is what makes it tamper-proof against us, is a configuration flip we have not shipped as default.

• planned

  DPDP §13, grievance redressal. The consent and correction machinery is live. A first-class grievance inbox is designed and not yet built.

None of these are hidden in a footnote. They are amber chips on a page any customer can open. We would rather lose a deal to a competitor's greener slide than win one on a claim we cannot defend in an audit.

Why this is the moat

Every serious agent will run on memory it did not write and cannot fully see. The memory layer that wins the enterprise will not be the one with the cleverest recall. It will be the one that can stand in front of a security review and answer, with receipts, what it stored, why, who saw it, whether it ever crossed a tenant boundary, and how to erase it on request without breaking its own audit trail.

We think that is a different category of product than a vector store with a memory API on top. We are building the one you can put on trial and have it hand you the evidence. As far as we know we are the first to publish a live clause-by-clause compliance matrix for agent memory, gaps and all.