I built an AI memory engine in 10 days, then needed a project to prove it works
10 hours ago
- #Verifiable Data
- #AI Memory
- #Merkle Proofs
- Merkle Proofs use RFC 6962-compliant SHA-256 Merkle trees; altering any atom changes the root hash, allowing AI clients to verify data integrity without trusting the server and saving 37% tokens compared to raw proofs.
- Markov Prediction employs a variable-order Markov chain to prefetch context by predicting future atom accesses, with weights decaying based on recency (0.5^(days/7)), achieving a 64% hit rate in production.
- Sub-ms Recall is achieved through LevelDB with JumpHash sharding across 4 independent Merkle trees, providing dedicated instances with zero contention and low latency (0.045ms p50, 1.2ms p99).
- MCP-Native support includes a streamable HTTP MCP transport compatible with Claude, Cowork, and other MCP clients, featuring OAuth2 + Bearer authentication and 25 MCP tools.
- Architecture ensures determinism and verifiability: atoms are hashed and committed to a Merkle tree upon deposit, sharded via JumpHash, with transitions recorded for Markov predictor updates and Merkle audit paths provided on reads.
- Built on Parametric Memory's own product, all internal operations (billing, health checks, etc.) are stored as Merkle-sealed atoms, demonstrating trustworthiness since March 2026.
- Pricing starts at $5/month with plans including Starter, Solo, Professional, Team, Enterprise Cloud, and Self-Hosted, featuring a 30-day money-back guarantee and no shared infrastructure.