EPCG

The question

What are the minimal typed interfaces through which an agent must access its own past — and can that decomposition be proven to generalize outside its training domain?

Overview

EPCG models persistent cognition in four categorical layers — semantic memory, relational memory, executive supervision, and an orchestration endofunctor. Its first falsifiable claim is already validated: flat retrieval (`retrieve(query) -> top-k`) is the wrong interface for agent memory. The Memory Access Model decomposes recall into four typed operations and beat flat retrieval by 50+ percentage points in blind generalization spikes (p < 10⁻⁷).

Approach

• 1.Memory Access Model v0: four typed operations selected autonomously by the agent — EpisodicAt(turn) for discrete events, StateAt(entity, dim, T) for point-in-time validity, StateHistory(entity, dim, range) for transitions, SemanticRetrieve(query) for open discovery.
• 2.Blind validation: N=200 tasks across two domains; flat retrieval scored 19.5% vs 70% for typed access (+50pp, p < 10⁻⁷) — zero LLM in the loop, isolating the architecture variable.
• 3.External audit on AMA-Bench: EPCG-T at 72.5% vs BM25 70.0% and LongContext 50.0%, with paired analysis and full provenance (frozen commits, raw JSONs, SHA256 manifests).
• 4.Category-theoretic composition: layers as categories, cross-layer bridges as functors, orchestration as a monad — making identity preservation provable rather than aspirational.

Open problems

Stated plainly, because pretending they're solved would be the opposite of research.

• Relational construction: current pilots run EPCG-T (temporal/selective); graph spreading between episodes remains untested territory.
• Closing the gap to dense baselines on semantic-only queries without losing the temporal guarantees.
• Governance of evolving memory: write-gates and reconciliation before self-edited memory compounds errors.