Chinese | English

dsh-ocr1-memory Status Overview

Quick links: README · Implementation Notes · Deployment Record · Benchmark · Test Report

Current Status

Item	Status
DSH plugin	✅ Installable; registers OCR1 optical tools and governance tools
Basic memory pipeline	✅ SoM rendering, age tiers, caching, OCR readback, verbatim Fetch
Locate-and-Transcribe	✅ Strict K-bit locator pipeline; requires a compatible trained model
Active recall	✅ Restores a fuzzy hit to vivid
Dynamic hit-heat decay	✅ Implemented; dynamicDecayEnabled is disabled by default
Per-turn context	✅ Enabled by default — injects L1/optical metadata; contextMode: snapshot keeps full-body snapshots
Visual embedding	✅ Saves real vectors when embeddingRetrieval is enabled; primary retrieval is off by default
Shared store	✅ Optional reload + atomic save
OCR server lifecycle	✅ Endpoint deduplication, cancellable startup, and cleanup of plugin-owned processes only
Maintenance and cancellation	✅ Maximum 8 entries per batch, namespace single-flight, disposal cancellation and drain
Governance integration	✅ Runtime skill, L1 index, pending distillation, turn/end maintenance, threshold reminders, synced with the same OCR1 store instance
Tests	✅ npm test 88/88 with the live backend available

Tools

• ocr1_mem_status
• ocr1_mem_store
• ocr1_mem_update
• ocr1_mem_retrieve
• ocr1_mem_list
• ocr1_mem_metrics
• ocr1_mem_calibrate
• ocr1_mem_forget
• ocr1_mem_render_test
• ocr1_mem_embed_test
• memory_read / memory_list / memory_write / memory_retrieve
• memory_search / memory_index / memory_pending / memory_accept
• memory_update / memory_archive / memory_rollback / memory_promote
• memory_stats / memory_maintain / memory_expand / memory_activate

Implemented Features

• Splits text by paragraph and length, then renders it as square images with SoM indices;
• Applies vivid → normal → fuzzy age decay, forcing regeneration and invalidating old OCR evidence when the resolution changes;
• Provides OCR readback and optional text/visual embedding recall; with requireOcr=false no-backend operation degrades safely, while strict mode reports failure;
• Provides request parsing, probabilistic selection, threshold/Top-K rules, and GBNF strict mode for the LoRA optical locator;
• Returns original verbatim text by segment index after Locate;
• Provides active recall, hit counts, bounded access history, and optional dynamic decay;
• Injects L1 index and optical metadata synchronously through systemPrompt.context() by default, with entry-count and character-count limits, without invoking the model inside the provider; snapshot mode still provides full-body snapshots;
• Provides a rendering cache, concurrency locks, an atomic manifest, multi-Agent sharing, and recovery from corrupted artifacts;
• Integrates the governance layer: namespaces, evidence, provenance, pending, archive, rollback, full-text search, and cross-namespace promotion; governance writes share the same OCR1 store instance.

Test Coverage

Suite	Count
Core store tests	12
Context tests	7
Complex isolation tests T1–T24	24
OCR HTTP	2
OCR server lifecycle	4
Embedding tests E1–E5	5
Locator tests L1–L8	8
Governance layer and cancellation signal tests	13
Integration wiring and automation tests	2
Rendering geometry RG1–RG2	2
Robustness M1–M9	9
Total	88

Real OCR, embedding, and rendering depend on backend services and the Python environment; with the current live backend all 88 cases pass, while eight live-backend cases skip when it is unavailable and the remaining tests still run.

Paper Reproduction Boundaries

The implementation provides engineering counterparts to OCR1/OCR-Memory: SoM, resolution decay, active recall, Locate-and-Transcribe, strict binary localization, optional embedding, dynamic decay, and context snapshots.

The following have not yet been fully reproduced:

1. DeepEncoder internal compression, per-layer visual-token counts, and internal tensor export;
2. Fully equivalent output from the official internal multimodal embedding;
3. Paper-scale training and complete main-table evaluations such as Mind2Web, AppWorld, and RULER;
4. Universal performance guarantees across arbitrary combinations of hardware, drivers, quantization, or service hosting.

These items require different levels of model-internal access, data/evaluation resources, or platform conditions; the current small-scale end-to-end validation should not be described as a reproduction of the paper's main tables. The current local evidence covers an end-to-end Windows CPU-only llama-server path; without a live backend, eight live-backend cases are skipped rather than treated as passing. For platform details and the validated model-conversion path, see DEPLOYMENT.md.