Now AvailableDedicated AI memory with cryptographic proofs. From $5/mo USD.View pricing →
Research

Published research

The ideas behind Parametric Memory are published, DOI-registered, and free to read. If you build on them, cite the DOI — and if you want to see the claims hold up at runtime, verify a sealed memory snapshot yourself or read the LongMemEval benchmark results.

Whitepaper · July 2026

Parametric Memory: A Cryptographically Verifiable, Predictive Memory Substrate for MCP-Capable AI Agents

The L2 cache for AI — verifiable, predictive agent memory behind one MCP interface

Abstract

AI agents reason brilliantly over their context window and forget everything when it closes. Retrieval-augmented generation restores recall, but a production agent also needs to know what it knew and when, what tends to come next, when two of its beliefs disagree, and how its facts relate. This paper describes a memory substrate delivering all four properties at once — Merkle-verifiable snapshots with RFC 6962-style consistency proofs, Markov-chain predictive prefetch, adaptive decay-based self-curation, and automatic contradiction detection — behind a single Model Context Protocol interface any MCP-capable AI can use.

Key contributions

  • Positions agent memory as an L2 cache between the context window and cold storage
  • Composes verifiability, prediction, self-curation and conflict-awareness in one substrate
  • Recall responses carry offline-verifiable cryptographic proofs of what was stored, and when

Keywords: AI memory · verifiable memory · Merkle proofs · RFC 6962 · Markov prediction · Model Context Protocol · agent memory substrate

Whitepaper · July 2026

The Self-Reinforcing Loop: Verifiable Evidence of Consensus for Distributed Agentic Systems

Signal → Evidence → Refinement on the Parametric Memory substrate

Abstract

A fleet of agents sharing a verifiable memory is a different kind of system from a single agent with one. This paper describes an architecture where fixed knowledge substrates give every agent identical versioned guidance, a state plane captures operational signals as append-only Merkle-versioned evidence, and an orchestrator refines a control plane of policies only when independent agents corroborate a signal and counterfactual replay shows the change would have improved outcomes. Consensus becomes a derived artifact any AI can recompute and verify from the evidence — asynchronously, offline, after the fact.

Key contributions

  • Replaces live consensus protocols with verifiable evidence-of-consensus
  • Graduated promotion: signals → corroborated evidence → counterfactually-gated refinements
  • Formal model: decay- and reliability-weighted support, correlation-corrected confidence, hysteresis thresholds

Keywords: multi-agent systems · distributed consensus · evidence · Merkle proofs · counterfactual evaluation · control plane · AI memory · Model Context Protocol

Both records are published under concept DOIs, which always resolve to the latest version of each paper. The Organization behind these works is described in machine-readable form on every page of this site.