Learning & Frameworks¶

How MemexLab turns long-form inputs into durable understanding that compounds. Three capabilities work together: extraction (get the knowledge in, with provenance), frameworks (interrogate it with mental models), and progressive learning (revisit, measure, and close gaps over time).

sources (books, reports, papers)
   │  memex-extract
   ▼
atomic items  ──  memex-frameworks (apply lenses, tag latticework)
   │
   ▼  memex-progress (coverage · gaps · spaced review · eval)
durable, compounding understanding

1. Extraction from reports & books¶

The memex-extract skill turns a long source into cited atomic items without losing provenance.

Chunk by chapter/section, preserving page or location references.
Extract candidates per chunk: claims, key concepts, notable quotes, decisions, open questions.
Provenance by default — every candidate carries source: (slug) and a page/location.
Dedupe before write — match candidates against existing items with BM25 and embeddings; merge into the strongest note rather than spawning near-duplicates.
Quotes are verbatim and attributed; inference is marked separately from source-stated fact.
Writes are dry-run; a human approves before apply.

For web sources, kepano/obsidian-skills' defuddle skill is a recommended companion — it turns a messy page into clean markdown that memex-extract then distills into cited items. See Obsidian Skills: comparison & interop.

Retrieval: BM25 + embeddings¶

Retrieval is hybrid and provider-agnostic. Deterministic BM25 gives exact-term recall and reproducibility; an embedding index (any provider — OpenAI, or a local model) gives semantic recall for paraphrased or conceptually-related material. Extracted items are indexed into both on write, so a query like "netting risk in deferred settlement" finds the relevant claim even when the source never used those words. The vault stays the source of truth; the indexes are a disposable acceleration layer, rebuildable from the markdown at any time.

2. Mental frameworks¶

Capture is cheap; judgment is the scarce asset. The memex-frameworks skill applies named lenses from the frameworks/ library — first-principles, inversion, second-order effects, base rates, incentives — as structured passes over a topic or item.

A lens produces assumptions, failure modes, questions, and framings — not invented facts.
Items are tagged by the framework used and the latticework problem they serve (problem-1 … problem-5, the five-problem meta-taxonomy).
The reasoning is cross-linked to the lens notes, so why a conclusion was reached stays inspectable.

This is how a raw claim becomes a thinking asset: the same fact, run through inversion and second-order effects, yields the risks and downstream consequences a flat summary misses.

3. Progressive learning¶

The memex-progress skill makes the vault improve over time instead of merely accumulating.

Coverage map — evergreen items per latticework problem and per domain; a readable signal of what you're thinking about and what you're neglecting.
Gap detection — under-served problems/domains, orphan notes, and stubs surface as the next things to learn.
Spaced revisiting — evergreen items become due for review on a recency × importance schedule; each pass strengthens, splits, or retires the note. (Review is surfaced, never applied unattended.)
Mastery signals — citations-in, links-in, eval scores, and contradiction flags per topic.
Learn-next — the highest-leverage gaps, and the sources that would close them.
Measured, not asserted — progress is tracked as eval-score and coverage deltas over time (see Benchmark Report and Observability).

How it compounds¶

Extraction fills the vault with cited claims; frameworks turn claims into judgment and tag them to the problems that matter; progressive learning revisits the weak spots and measures whether quality is actually rising. Each loop deepens the corpus and improves the next answer — the opposite of re-deriving from scratch every session.

See it end to end¶

A complete worked example runs the pipeline over a real public source — Vannevar Bush's As We May Think (1945, the essay that named the memex): ingest → extract → a mental-model lens → progress, with cited artifacts that pass validate_vault.py.

← Engineering & design · Operating System · Taxonomy