ARMS & HAT Memory Lab

Lab Status

This active research lab collects questions and implementation references for HAT and ARMS. The four visualization and analysis interfaces listed here are planned; they are not embedded, working experiments on this page.

Editorial update — September 22, 2026: This record now follows the scope and limitations in the corrected papers. No new experiment or independent reproduction accompanies this update.

Read the Research

HAT: Retrieval Through Known Structure

HAT organizes vectors using a supplied session–document–chunk hierarchy and searches through container centroids. The reported synthetic benchmarks favor that known structure. Recall depends on the data and search settings; it is not a universal guarantee. HNSW had lower query latency in the reported comparison, while HAT had higher recall and faster construction.

The HAT paper record separates the small hierarchical benchmark, scale series and larger-corpus retrieval example. It also documents unresolved figure accounting and distinguishes retrieval recall from downstream LLM answer accuracy. Use that record for the measured values, conditions and original paper link.

ARMS: Proposed Persistent State Memory

ARMS proposes a spatial architecture for storing and retrieving model-state payloads. The separate ARM prototype uses a learned projection and a coordinate index. A retrieved record is not evidence of exact tensor reconstruction, and the reported compression ratio lacks the complete storage accounting needed to establish lossless compression.

The ARMS paper record distinguishes the proposal from the early prototype and explains the missing storage, reconstruction and model-integration evidence. Its results do not establish unlimited context or general lossless memory compression.

Available Project Links

• HAT source: experimental hierarchical index implementation and benchmark notes.
• ARMS source: early memory architecture and adapter implementation.
• Infinite Context source: related project code; follow its documentation for current setup and supported interfaces.
• Infinite Context on Hugging Face: external hosted project page. The page was reachable during this revision; its runtime and outputs were not validated here. “Infinite Context” is the project name, not a demonstrated unlimited-context capability.
• Generic 3D visualization: a separate browser visualization demo. It does not display a measured HAT tree or validate ARMS memory behavior.

Installation and APIs can change. Use the linked repositories for package names, supported commands and current examples. Source availability does not establish independent reproduction of a paper’s results.

Planned Interfaces

HAT Tree Visualization

A proposed view would expose the global root, sessions, documents, chunks and centroids. It should make the supplied hierarchy visible and distinguish real benchmark data from illustrative examples.

Coordinate Explorer

A proposed explorer would compare learned coordinates and projected views of them. A t-SNE or UMAP plot may help inspect a dataset, but visible clustering would not by itself establish accurate retrieval, lossless storage or downstream task quality.

Memory Retrieval Interface

A proposed query view would show the candidates examined at each level, similarity scores, beam settings and returned records. Useful comparisons would include ground-truth neighbors, baseline settings, query latency and failures on less structured data.

Storage Accounting Analyzer

A proposed analyzer would report original tensor shapes and precision, stored payload bytes, coordinate bytes, metadata, encoder weights and any excluded or amortized costs. Reconstruction error and downstream task results would be reported separately from the storage ratio.

Research Questions

• When does a supplied conversation hierarchy improve retrieval, and when do incorrect or weak groupings make it worse?
• How do HAT and baseline indexes compare at matched recall, with documented settings and measured latency?
• Which stored states can be reused across queries or sessions without changing a model’s answers in unacceptable ways?
• What is the total storage and inference cost once payloads, coordinates, metadata and projection models are included?
• Does retrieving the expected record lead to better downstream answers on held-out tasks?

Design Analogy

The projects use human episodic memory as a design analogy: a limited working context can consult an external store. This analogy is not a validated model of the hippocampus or evidence that an AI system has human-like memory.

Continue with the HAT overview, ARMS overview, or the research archive for the current public records and their limitations.