RecallForge v2.2.0 — Multimodal Material Intelligence

RecallForge v2.2.0 — Multimodal Material Intelligence

RecallForge can now begin where real exam preparation usually begins: lecture slides, PDFs, scans, images, and past papers.

Highlights

• Multimodal Material Intelligence: page/slide-level routing from native text, host vision, and optional local OCR, fused into a source-grounded StudyDocument IR with zero silent page drops.
• Source-grounded adaptive review: every reconstructed concept, past-paper question, and uncertain formula keeps a source anchor; knowledge structure → diagnostic recall → weakness diagnosis → corrective teaching → follow-up practice.
• Better first-use experience: 30-second $recallforge self-test, $recallforge multimodal-test, $recallforge inspect-materials, bilingual guides, and clean-room verified ZIP/Plugin packages.

What is included

• PPT/PPTX: slide order, titles, text boxes, native tables, speaker notes, image/chart blocks, spatial layout, source anchors.
• Digital PDF: per-page native text and layout blocks with selective visual verification.
• Scanned PDF and images: per-page scan detection, host-vision-first routing, optional local OCR fallback.
• Past papers: question/options/score/subquestion structure; printed answers stay separate from user or unknown annotations.
• Formulas and tables: raw/interpreted formula pairs with confidence; headers/rows/cells preserved; conflicts stay uncertain rather than silently rewritten.
• Diagram-heavy subjects: chemical structures and biological figures remain visual concepts; visual emphasis and explicit teacher wording are treated as source-bounded evidence.

Verification

• Full automated suite: 215 passed locally; GitHub Actions passes on Ubuntu 24.04, Windows, and macOS.
• Official Skill and Plugin validators pass; clean-room ZIP/Plugin install tests pass.
• Real Codex host E2E (Codex 0.147.0, Windows 11): /skills discovery, text self-test, multimodal self-test, functional test, and a real material review (16 lecture decks + 30-page scanned past-paper PDF; 1047 slide/page units + 30 exam pages) all passed. Evidence: verification/host-verification-template.json.
• Local OCR is optional and experimental. Reference results (CPU): Tesseract 5.5.0 mean CER 0.2578; RapidOCR 1.2.3 mean CER 0.2492 across 10 self-authored fixtures. Neither engine is treated as verified understanding; host vision remains the recommended path.

Known limitations

• Local OCR cannot verify formulas, diagrams, table layout, or exam structure alone.
• Scanned pages with rotation, stitching, handwriting, and legacy PPT visual rendering remain recognition boundaries that are reported rather than silently accepted.
• Host vision quality depends on the host/model and is not claimed as universally verified.

Installation

Most users should download: recallforge-skill-v2.2.0.zip

1. Extract the zip and copy the recallforge folder to %USERPROFILE%\.agents\skills (Windows) or ~/.agents/skills (macOS/Linux).
2. Start a new Codex turn.
3. Run $recallforge self-test and confirm Status: READY.
4. Attach course materials and run $recallforge inspect-materials before starting review.

Plugin users: recallforge-plugin-v2.2.0.zip provides the skills-only Plugin package.

中文摘要

RecallForge v2.2.0 是本项目第一个“多模态资料智能”正式版本。它不再要求用户先把 PPT、扫描教材、往年试卷整理成 Markdown，而是从真实原始资料开始：逐页/逐幻灯片路由、原生文字、宿主视觉与可选本地 OCR 融合，并始终保留来源锚点和不确定性标记，再进入知识重构、主动回忆、薄弱点诊断、纠正教学与模拟考试。

最常用下载文件：recallforge-skill-v2.2.0.zip

Assets

• recallforge-skill-v2.2.0.zip — recommended manual / Codex Skill installation
• recallforge-skill-v2.2.0.tar.gz — macOS/Linux archive
• recallforge-plugin-v2.2.0.zip — skills-only OpenAI plugin adapter
• SHA256SUMS.txt — asset checksums

Upgrade notes

v2.1.3 was an internal tagged packaging correction and did not receive a formal GitHub Release. Users upgrading from v2.1.2 can replace only the recallforge Skill folder and rerun $recallforge self-test.