Now AvailableDedicated AI memory with cryptographic proofs. From $3/mo.View pricing →

What is Parametric Memory?

Parametric Memory is a dedicated memory service for AI agents. It gives Claude — and any other MCP-compatible AI — a persistent memory that survives between sessions, can be verified with cryptographic proofs, and gets smarter the more you use it.

This is not a vector store. It is not a notes file. It is structured, tamper-evident memory with a Markov prediction layer that learns your patterns and pre-fetches what you're about to need.

The three guarantees

GuaranteeHow it works
PersistentAtoms (named facts) survive indefinitely across sessions. Your AI picks up exactly where it left off.
VerifiableEvery stored atom is hashed into an RFC 6962 Merkle tree. You can verify any recalled fact was actually stored — no hallucinated memory.
PredictiveA PPM Markov model learns from your recall patterns. It achieves a 64% average hit rate pre-fetching atoms you're about to need, reducing latency to near zero.

How memory actually works

Everything stored in Parametric Memory is an atom — a key/value pair with a namespaced key like v1.fact.my_database_url.

When you store an atom:

  1. The value is hashed (SHA-256) and inserted into a Merkle tree for your shard
  2. The Markov model records which atoms were accessed before and after this one
  3. Future sessions that touch nearby keys will have this atom pre-fetched automatically

When you recall atoms:

  1. The system returns matching atoms plus a Merkle proof for each one
  2. You (or your AI) can verify the proof against the tree root — confirming the fact was not altered
  3. The Markov model notes this recall and strengthens the arc weights for related atoms

Your dedicated instance

Every paid subscriber gets their own isolated memory substrate — a containerised instance running on our infrastructure. Your atoms never share compute or storage with another customer.

  • Starter / Solo / Professional plans — shared host, isolated container per customer
  • Team plan — dedicated DigitalOcean droplet, isolated at the machine level

Your instance has its own domain, SSL certificate, and MCP endpoint.

What it's not

  • Not a file store — atoms are small named facts, not binary blobs
  • Not a chat history — it stores structured knowledge, not conversation logs
  • Not a RAG pipeline — there are no embeddings or cosine similarity; recall is exact key lookup or semantic search over structured fields
  • Not a database — it is an append-only memory layer optimised for AI agent patterns

Next steps

What is Parametric Memory? | Parametric Memory