Engram
GitHub ↗
macOS · indexes the knowledge · concept graph · chat
Engram turns Microsoft OneNote courses (exported as PDF) into an Obsidian-native Markdown vault on your Mac. It's an index, not a transcription: each page becomes a concise, searchable summary of its key ideas and formulas, the original right below. Related notes link into a multi-layer concept graph — and now you can chat with it, grounded in your notes and cited.
A research prototype by Kaiwen Bian
You already write notes on an iPad with Apple Pencil in OneNote. Big notebooks can't be exported whole, so you export one section group at a time (e.g. one course) to PDF. Engram groups several such PDFs into a single domain — treated as one combined knowledge base — and turns it into something you can actually think with.
Domain → Section group → Section → Note gives you the tree. A holistic pass then distills the whole domain into a multi-layer concept map — Concepts → Themes → Areas — where nodes cluster by parent and connect by shared notes (solid) and model knowledge (dashed). The macOS app renders each layer as an interactive, force-directed graph — switch layers, drag a node, or click one to see the notes it covers.
GraphEngine (repulsion · springs · gravity · clustering). Click a node.
Each note keeps its original page as a faithful per-note PDF leaf; above it sits a concise, searchable index entry — the page's key ideas and all its formulas — always one scroll from the original. Engram indexes the knowledge, not the page: nothing is invented, illegible parts are flagged, and papers / homeworks / long code are skipped.
Snap the tail down, then slide the front foot up the board to level it out.
![[Pop & Timing.pdf]] # the raw leaf, embedded<Domain>/<SectionGroup>/<Section>/<Title>.md beside .pdf.## Related wikilinks — even across section groups.Once the concept graph exists, you can chat with your notes. Context is pulled deterministically from the graph and your summaries — no embeddings — and every answer is grounded and cited (the notes it used become clickable sources), with anything beyond your notes clearly marked. Two modes: Query a single course or concept, or Synthesise a whole field.
How does value iteration converge, and why?
Thinking…
Value iteration converges because the Bellman operator is a γ-contraction in the sup-norm:
‖ℬV − ℬV′‖∞ ≤ γ‖V − V′‖∞, γ < 1
so the value estimates form a Cauchy sequence with a unique fixed point V* — geometric convergence from any start.
▸ Beyond your notes — next: asynchronous / prioritized sweeping, sample-based VI.
[[sources]].Built for a personal knowledge base you actually trust — local, faithful, and reproducible.
Pick the reader in Settings: Claude Code (vision, via your login — best quality, default) or a local Qwen2.5-VL 7B/32B, fully on-device. The concept-map & cross-link reasoning runs through Claude Code too — all from a dropdown, no code.
Engram distills each note to a searchable digest of its key ideas and all formulas — so you can find and reconnect, then jump to the original page (the source of truth, embedded in every note). It filters out noise — attached papers, homeworks, long code — and never invents; illegible parts are flagged.
Beyond folders, a holistic pass builds a hierarchy of concept graphs — Concepts → Themes → Areas — that span section groups. Nodes cluster by parent; links are grounded shared-note edges plus dashed knowledge edges. Switch layers as tabs in the app.
Ask your knowledge base. Query a course or a concept, or synthesise a whole field. Context is pulled from the graph + your notes; answers are grounded and cited (clickable sources), with anything beyond your notes clearly marked. Your choice of Claude model.
Re-add one course and only its changed or new notes are re-processed, keyed on a rendered-page fingerprint — the rest of the domain is untouched. An unchanged export is a true no-op (the model never even loads), and every run reports what changed.
Plain Markdown on disk — frontmatter, [[wikilinks]], embedded PDFs, section · section-group · domain overviews. Open the folder in Obsidian, or paste big slices into a long-context LLM (Karpathy's "second brain").
A SwiftUI app to add/replace a section-group PDF (with a change report), browse notes by section group → section, read each summary with LaTeX & code rendered beside its original page — and explore the multi-layer concept graph.
Python 3.11+ and uv on macOS 14+ (Apple Silicon). The first run downloads the local model (~5 GB).
# 1 · set up the environment (add --extra local for the on-device model)
uv sync --extra dev --extra local
# 2 · create a domain, then add a section-group PDF (one per course)
uv run python -m engram domain create "Data Science"
uv run python -m engram add "Data Science" "CSE 234.pdf"
# 3 · cross-link the whole domain (links span section groups)
uv run python -m engram link "Data Science"
# 4 · build the multi-layer concept map → the graph
uv run python -m engram concepts "Data Science" --layers 3
uv run python -m engram statusOutput lands in ~/Engram/<Domain>/<SectionGroup>/<Section>/<Title>.{md,pdf} — open ~/Engram in Obsidian.
Prefer a GUI? Build the native app:
cd macapp && ./build_app.sh && open Engram.appEngram's whole ethos is honesty over polish. This is a working v0.2, run end-to-end on real multi-course domains (and v0.1 validated on a known notebook of 18 notes, 73 links) — but it's early, and several things are heuristic or not yet field-validated.