What question did this study set out to answer?

The aim is to clarify the AAMT architecture and its distinctive components, separating it from existing frameworks.

July 11, 2026Open Access

The AAMT Platform Architecture: Construction, Runtime, Ecosystem Integration, and Comparative Position

Key Points

The aim is to clarify the AAMT architecture and its distinctive components, separating it from existing frameworks.
Introduced a four-dimensional coordinate space (TERA) with a 16-orthant tessellation (Meji) and a 256-cell refinement (Odu).
Developed a memory-mapped binary format for seamless cross-platform functionality.
Outlined various runtime mechanisms, including retrieval strategies and redundancy schemes.
The AAMT architecture operates independently of model weights, enhancing its functional claims.
It was implemented in six real applications, demonstrating its practical utility and integration.
Comparative analysis shows distinct capabilities compared to transformer stacks, vector databases, and RAG systems.

Abstract

The AAMT stack is usually described one paper at a time — a lattice here, a substrate there, an escape-steering rule somewhere else — and the throughline is easy to lose. This paper states it directly: AsAManThinks runs a geometric memory layer beside conventional language-model inference, not instead of it. A 4-dimensional coordinate space (TERA), an exact 16-orthant tessellation of it (Meji) with a further 256-cell refinement (Odu), and a small number of provably deterministic operators (a chiral tie-break Ox, a product Fisher–Rao metric, a Householder reflection) form a mathematical layer that is deliberately independent of any model weights. A memory-mapped binary format (.aamt, WP-23) makes that geometry a zero-parse, cross-platform artifact — the same bytes open identically from a native CLI, from Python via ctypes, from a browser via WebAssembly, and from Unity via P/Invoke, with byte-identical routing verified across all four. A small, orthogonal set of runtime mechanisms sits on top of the geometry: coarse-then-fine retrieval (WP-20), rolling output eviction for unbounded generation (WP-21), session-level entropy monitoring with cross-session escape steering (WP-22, WP-23 §9), and a geometric redundancy scheme built from a single exact reflection rather than a generic checksum (WP-23 §10). We describe how this threads into six real, running consumers across the platform — the desktop training/inference stack, the DreamOS Unity client, the public math visualizer, the Oracle divination app, the InnerSpeech translation engine, and the marketing site's production deployment — and we state plainly where the architecture's claims are measured and where they are still aspirational. We then compare the resulting shape against three architectures it is easy to conflate it with — token-sequence-only transformer stacks, conventional vector databases, and generic RAG pipelines — and show the difference is not cosmetic: it is a difference in what the system is willing to claim it knows without running a model at all.

Read Full Paperexternally

Perguntar à IA

Bookmark

View Full Paper