Parametric Memory (MMPM)

Parametric Memory (MMPM) is the L2 cache for AI agents — persistent, verifiable memory that keeps the right context warm before your agent asks.

Every memory is written to a SHA-256 Merkle tree, so your AI can prove nothing was altered. A Markov chain prefetches what's likely next — 64% hit rate in production. Recall is 0.022ms p50. 76.6% on LongMemEval-S with zero setup, 83.0% with typed ingest. It's an MCP server: 11 tools, one config block, works with Claude, Claude Code, Cursor, Cline and any MCP client.

On this channel:
- L2 memory demos — end to end, in real time: empty substrate to an agent that remembers
- Agentic workflow design — building agents around a memory tier instead of a bigger prompt
- AI workflow optimisation — cutting tokens, latency and re-reading by recalling instead
- The science of a memory substrate — Merkle proofs, Markov prefetch, decay, the edge codec

For engineers shipping agents that have to remember. From $5/mo — parametric-memory.dev