Gespeichert in:
| 1. Verfasser: | |
|---|---|
| Format: | Recurso digital |
| Sprache: | |
| Veröffentlicht: |
Zenodo
2026
|
| Schlagworte: | |
| Online-Zugang: | https://doi.org/10.5281/zenodo.20062417 |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
Inhaltsangabe:
- <p>Long-term AI memory systems face a fundamental architectural choice: store by similarity (embeddings) or store by structure (knowledge graphs with explicit relations). Recent geometric analysis of pre-trained embedding models shows that production-scale embedding spaces concentrate around d_eff ≈ 16, and that storage-by-similarity inherits canonical interference-driven failure modes — power-law forgetting, false-memory recall at semantic-cluster centroids, and consolidation-via-centroid-merging that incurs roughly 4× backward interference for 62.5% compression — as mathematical consequences of low-effective-dimensional storage geometry (Barman et al. 2026). These are not implementation defects; they are structural properties of similarity-based memory at scale.</p> <p>We present NSP (Narrative State Protocol), a structural cognitive engine that retrieves by graph traversal over typed edges with explicit provenance, never by cosine similarity. Its four-layer organization — Universal Substrate (Layer 0, domain-agnostic), Cognitive Scaffold (Layer 1), Encoding Adapter (Layer 2, the only domain-specific code), and Expression Adapter (Layer 3) — admits four immunity sub-claims by structural design: storage immunity, consolidation immunity, false-memory immunity, and partial-staleness immunity at the storage/retrieval layer (with LLM-processing-layer interference managed via injection budgets, disclosed honestly).</p> <p>We exercise the same engine across three radically different domains — combinatorial matrix research (Hadamard 668, with Paper-2-cited Williamson 65 cross-problem transfer), software engineering (deployed self-instrumentation as nsp-coding), and music composition (mgen ballad-melody generation, V4→V7c arc) — and report four experiments: (i) corpus-size-stability of retrieval precision against an embedding baseline (variance-stability difference rather than the bAbI-style precision-decline curve Barman et al. observed — the so-called "HIDE" shape — on three corpora that differ in shape from bAbI); (ii) knowledge-driven generative improvement under a single human listener oracle, with the V7→V7b transition serving as a causal-control before/after pair tying axiom application to listener-perceivable change; (iii) per-category cross-surface axiom transfer (orchestration / mix / melody) across ten transfer surfaces beyond the development surface (five within two design clusters plus five genre-anchored α-prime surfaces); (iv) a three-domain cognitive-profile snapshot showing that the same engine, deployed across math, coding, and music with the integration-depth caveat below, produces three measurably different cognitive profiles — sparse-and-deep KG for math, bushy-and-shallow KG for coding, layered-and-narrow catalog for music — through the interaction product of the universal substrate and each domain's truth-condition signal.</p> <p>Structural cognitive scaffolding is sufficient to span three radically different domains under one engine without per-domain reengineering, with the integration-depth caveat that the math and music instances exercise the full cognitive-engine stack while the coding instance currently exercises the cusp-catastrophe belief dynamics but not yet the axiom-application loop (§8.1 L4 + §8.2 F7 detail this scope and the planned remedy), and it does so under both formal-verification and subjective-listener regimes.</p>