memorise

Memorise them All

Turn any pile of documents into local, token-free graph memory for Claude, and recall from it cheaply. All conversion and digestion runs on the user's machine (MarkItDown, plus optional Tesseract OCR) — fully deterministic and model-free, no LLM/Ollama/GPU. Claude only ever issues a small tool call and gets back compact metadata or a tiny relevant slice — never whole documents.

When to use

• The user wants to ingest/digest/memorise files, attachments, or a whole folder.
• The user asks a question that should be answered from their own documents.
• The user wants a knowledge graph of their documents.

How it works (the pipeline)

1. Convert — every attachment → Markdown locally (PDF/Office/HTML via MarkItDown, incl. legacy binary .doc/.ppt/.xls via optional LibreOffice; scanned images via optional Tesseract OCR; archives unpacked safely). Legacy Bengali typed in Bijoy/SutonnyMJ ANSI fonts is auto-upgraded to Unicode (font-aware, so mixed English+Bengali documents convert cleanly).
2. Segment → Extract — structure-aware chunks; rule-based, deterministic extraction of entities, relations, and atomic facts (no LLM/model).
3. Graph + themes — a knowledge graph with community-detected themes.
4. Layered memory — a global synopsis, per-theme summaries, per-document notes, and graph.json. Recall ranks with model-free BM25 (Bengali-aware) and returns a tiny cited slice, declining off-topic queries.

Tools

• digest(paths, project?, reset?) — build/refresh memory (reset: true rebuilds from scratch). Returns metadata only.
• convert(paths, out_dir?, project?) — convert files/dirs/globs to Markdown locally (legacy Bengali/SutonnyMJ → Unicode); writes .md files to out_dir (default markdown_converted/ beside the input). Token-free. Use when the user just wants Markdown, not a digest.
• recall(query, project?, k?) — return a small, citable slice of memory.
• memory_overview(project?) — synopsis + themes.
• export_memory(dest, project?) — export portable Markdown files.
• list_digestible(directory) — list convertible files (paths/sizes only).
• forget(project?) — delete a project's memory (graph, converted Markdown, summaries/notes). Irreversible.
• memory_status() — local stack health.

Rules

• Never read the converted documents back into the conversation — that defeats the token-free design. Trust the metadata and recall slices.
• Always cite source document names from recall hits when answering.
• If a project has no memory yet, run digest first.
• Use a project name to keep separate, reusable memories per body of work.