KAL: Connecting Knowledge Graphs to Geometric Memory Systems

Abstract

Prior work in this series has developed Geometric Memory Systems (GMS) along three dimensions: a benchmark for measuring structured retrieval, an evaluation framework showing that LLM-as-judge is unreliable on structured tasks, and a multi-agent architecture in which GMS serves as both persistent memory and governance mechanism. GMS itself supports multiple downstream consumer modes—verification primitive, agent memory, validation harness—but across all of them the knowledge graph that GMS operates over is assumed to exist. This paper addresses the assumption directly: how should a knowledge graph be designed and exposed so that GMS can consume it cleanly? We introduce the Knowledge Adapter Layer (KAL)—the knowledge plug-in for GMS, and a design specification for the data plane that sits between a knowledge graph (relational store, native graph database, or RDF endpoint) and GMS, regardless of how GMS is used downstream.

KAL has three design components. The adapter protocol is a small typed Python interface that any knowledge graph can implement: it groups operations into reads, writes, and similarity search, and pairs them with a capability descriptor that says which subset of the surface a given backend actually supports. The data crossing the protocol boundary—typed nodes, triples, literals, queries, and query results—carries the per-triple metadata GMS consumes (confidence, status, per-engine scores, evaluator identity) as first-class fields rather than opaque payloads. The federation router takes multiple adapters and presents them as a single graph to the caller: it fans queries out concurrently, merges results under a backend-independent content-hash dedup rule that handles both node-object and literal-object triples, enforces per-adapter timeouts so one slow source cannot stall the rest, and surfaces partial failures as structured error data so callers can distinguish “adapter returned nothing” from “adapter timed out.” The schema discipline is the contract a customer’s knowledge graph must satisfy in order to participate: typed entities with stable identifiers, a controlled predicate vocabulary, edge-attached provenance and verification metadata, and an external-ID registry that lets the same logical entity be addressed through Neo4j node IDs, Postgres UUIDs, or SPARQL URIs without identity collisions across stores. The protocol is symmetric under a second federation axis: the same triple can carry verification metadata from multiple GMS instances—a parent agent and its sub-agents, or multiple departmental GMSs within one organisation—and the federation primitives the protocol already exposes serve both axes without surface extension. A reference design grounds the specification: a Postgres adapter against the existing Knowlytix relational schema. The paper closes with KG autotuning as a self-improving consumer mode in which GMS’s own geometric signals propose typed mutations to the KG through the same protocol surface—turning KAL-readiness from a one-time onboarding contract into a property the deployment maintains under drift, and letting GMS land as a self-improving knowledge system rather than as a verifier-in-isolation. KAL is released as knowlytix-kal under Apache-2.0.

A provisional patent application covering the Knowledge Adapter Layer was filed by Knowlytix in May 2026. Patent pending.