Now AvailableDedicated AI memory with cryptographic proofs. From $3/mo.View pricing →
Back to blog
launchengineeringmerkle-proofs

Introducing Parametric Memory: Verifiable AI Memory with Merkle Proofs

·Glen Osborne

Introducing Parametric Memory

Today we're launching Parametric Memory — a dedicated memory system for AI agents that makes every recalled fact cryptographically verifiable.

The problem with AI memory today

Most AI agents are stateless. Each conversation starts from scratch. Those that aren't stateless typically rely on vector databases — which are probabilistic by nature and offer no way to verify that a retrieved fact wasn't corrupted or hallucinated.

We think memory should be more like a database than a dream.

What we built

Parametric Memory combines three primitives that don't usually appear together:

1. Merkle-hashed atoms

Every piece of information stored is an atom — a named, versioned string with a SHA-256 hash. Atoms are inserted into an RFC 6962 Merkle tree. When you recall an atom, you receive a proof alongside the value:

const { atom, proof } = await memory.recall("user preferences");
const valid = await memory.verify(atom, proof);
// → true (or throw if tampered)

No trust required. You can verify the recall yourself, client-side.

2. PPM Markov prediction

We trained a Prediction by Partial Matching (PPM) model on recall sequences. By observing which atoms you access in which order, the model pre-fetches what you're likely to need next — before you ask.

In internal testing against real agent sessions, we're seeing a 64% hit rate on predicted-next atoms. That means 64% of the time, the atom you need is already in your cache when you ask for it.

3. MCP-native transport

Parametric Memory ships with a Model Context Protocol (MCP) HTTP server alongside its REST API. Connect any MCP-compatible agent — Claude, Cursor, Cline — and it gains persistent memory immediately, with no SDK required.

Performance numbers

We obsess over latency. Here are real numbers from our production deployment:

MetricValue
Recall p500.045ms
Recall p991.2ms
Batch recall (50 atoms)0.8ms
Markov prediction hit rate64%

What's next

We're launching with Starter, Solo, Professional, and Team tiers. Plans start at $3/month (Starter).

Get started now — plans from $3/month.

The roadmap includes consistency proof auditing, multi-tenant isolation, and a hosted knowledge graph explorer. If you have requirements we haven't covered, reach out — we're building this in the open.

Introducing Parametric Memory: Verifiable AI Memory with Merkle Proofs | Parametric Memory