Cryptographic Trust Layer for AI Memory

Prove what your AI knew, and when it knew it.

Attested Memory gives AI systems a tamper-evident, cryptographically verifiable record of memory state at decision time — provable to auditors and regulators without exposing the underlying data.

Read the Architecture
SOC 2 in progressBuilt for regulated industriesSelf-hostable
Decision Receipt
VERIFIED
decision_id
dcn_01HZQK3V8N7X2P4R6M9T5B
timestamp
2026-07-12T14:22:07.481Z
root_hash
0x9f4c…a71d
model
internal/policy-v3
entries
1,284 chained writes
anchor
ots://btc/863,201
sha256:c1d4b2…f0e3a7 · leaf#00412
sha256:a83e19…7b2c4d · leaf#00413
sha256:5f7d0c…19ab8e · leaf#00414

AI memory today is a black box.

Vector databases, RAG pipelines, and knowledge graphs give AI systems recall — but not proof. Records can be silently edited, deleted, or contradicted, and there is no way to reconstruct what an AI actually knew when it made a specific decision.

  • No audit trail

    You can inspect the current state, but you can't prove what memory looked like at the moment of a decision.

  • No tamper detection

    Vector stores and knowledge graphs can be silently edited — after the fact, no one can tell.

  • No selective disclosure

    Sharing evidence with a regulator means exposing the entire memory, or asking them to trust you.

The Solution

A cryptographic ledger for AI memory.

Every memory write is committed to a tamper-evident hash chain, producing verifiable proof of what was known — without needing to trust the database operator.

01

Tamper-Evident by Construction

Every write is hash-chained to its predecessor, so any alteration breaks verification instantly and provably.

02

Selective Disclosure

Prove a specific fact existed in memory without revealing anything else — designed for regulators, not exposure.

03

Independently Verifiable

Anyone with a receipt can verify it against a public root. No trust in our servers required.

From memory write to verifiable proof.

Memory Write
Hash Chain
Merkle Root
Verifiable Receipt
  1. 1

    Memory is written

    Any write to your vector store or knowledge graph passes through the Attested layer.

  2. 2

    Entry is hashed and chained

    The write is content-hashed and appended to a per-tenant hash chain.

  3. 3

    Root is committed and anchored

    Chain roots are periodically anchored externally, creating an immutable timestamp.

  4. 4

    Proof is generated on demand

    Any decision can produce a compact receipt provable against the anchored root.

Built for systems that need to prove, not just remember.

Hash-Chained Writes

Every write inherits from the last, so any tamper breaks the chain.

Merkle Proof Generation

Compact proofs let you verify a single entry without the whole tree.

Decision Receipts

Each AI decision emits a signed receipt tying it to a memory root.

External Anchoring

Roots are anchored to public timestamps to remove operator trust.

Redaction-Aware Disclosure

Prove a fact was known without revealing surrounding memory.

Framework-Agnostic API

Wraps any vector DB or memory store — no lock-in, no rewrites.

Security and compliance, by design.

Self-hostable

Deploy inside your own VPC or air-gapped environment.

Zero-knowledge option

Prove facts without exposing the memory that produced them.

Audit-ready reports

Exportable evidence bundles for internal and external audits.

BYOK / your infra

Bring your own keys, your own storage, your own boundaries.

Purpose-built for regulated industries.

FinTechHealthcareInsuranceGovernmentEnterprise AI

Deploy in your jurisdiction of choice. Self-hosted, region-pinned, or fully air-gapped — Attested Memory never forces data across borders.

Add it to any memory system in an afternoon.

Attested Memory sits in front of your existing vector database or memory store as a thin middleware layer — framework-agnostic, no data migration, no lock-in.

  • Works with any vector database
  • REST and SDK access
  • Deploy self-hosted or managed
example.ts
import { attested } from "attested-memory";

const memory = attested.wrap(vectorStore, { tenant: "acme" });

await memory.write({ key: "policy.v3", value: doc });

const receipt = await memory.receipt(decisionId);
// -> { root, proof, timestamp, signature }

See a proof, verified.

Recent Receipts
Receipt
rcpt_9f4c…a71d
Proof Verified
decision
policy.v3 · underwrite
timestamp
2026-07-12T14:22:07Z
Proof Path
root0x9f4c…a71d
leaf#00412c1d4b2…f0e3a7
leaf#00413a83e19…7b2c4d
leaf#004145f7d0c…19ab8e
leaf#00415e2b641…0c73fa

Simple pricing, built to scale with trust.

Developer

For solo hackers and proof-of-concept builds

$0/mo
  • 5,000 receipts / month included
  • Hard stop at limit — no surprise billing
  • 14-day log retention
  • Local Firestore verifier only

Growth

For startups running live agents in production

$49/mo
  • 25,000 receipts / month included
  • $0.40 per 1,000 receipts overage
  • 90-day log retention
  • Programmatic verification SDK
  • Email support
Most Popular

Business

For growing teams scaling past prototype

$199/mo
  • 100,000 receipts / month included
  • $0.40 per 1,000 receipts overage
  • 180-day log retention
  • Programmatic verification SDK
  • Priority support

Enterprise

For fintech, healthtech, and regulated enterprise

Custom
  • Unlimited receipts, volume-based pricing
  • Unlimited log retention
  • External anchoring (independent, publicly verifiable timestamping)
  • SOC 2 / HIPAA-ready audit report templates
  • SSO, dedicated support, custom SLAs

All plans include tamper-evident hash-chained storage and Merkle proof generation. Enterprise-tier features roll out progressively — early customers get direct input on the roadmap.

Trusted by teams who can't afford to guess.

We used to argue with auditors about what our model saw. Now we hand them a receipt and the conversation is over.
Placeholder Name
Head of AI Risk · Regional Bank (placeholder)
Selective disclosure is the feature we didn't know we needed. It let us satisfy the regulator without exposing the underlying case file.
Placeholder Name
Compliance Lead · Health Insurer (placeholder)
It slotted in front of our existing vector store in a day. Two weeks later we had an audit-grade trail for every decision.
Placeholder Name
Platform Engineer · Enterprise SaaS (placeholder)

Questions, answered.

No. Attested Memory sits in front of your existing vector database or memory store as a middleware layer. Your data stays where it is; we add a verifiable trail on top.

An audit log records events but can be edited or contradicted by whoever controls it. Attested Memory produces cryptographic receipts that anyone can independently verify against an anchored root.

Yes. The Enterprise tier runs fully self-hosted or air-gapped, with your own keys, storage, and anchoring backends.

You can prove that a specific fact existed in memory at a specific time without revealing the surrounding memory. Useful for regulators, auditors, and customers.

Yes. The system is designed to run without external network dependencies, with optional local anchoring for fully offline environments.

By the volume of memory writes committed to the chain per month, plus optional add-ons for external anchoring frequency and support tier.

Give your AI memory something to prove.

Add cryptographic proof to any memory system. Ship auditable AI in an afternoon.

No credit card required · Deploy in an afternoon