Skip to content

MemexLab in One Page

What it is. A markdown-native governed memory layer for agents that turns reading, thinking, and daily capture into durable, queryable, linkable knowledge — with an LLM-powered engine that helps you compile sources into atomic notes and generate essays, decks, and answers from them.

Evolving toward: The Memex Mark 1 Operating Core — a sovereign operating system: Graph + Evidence + Belief + Action + Governance. The Mark 1 nine-layer architecture describes where the system is headed; this doc describes what ships today.

What makes it different. The filesystem is the database. Plain markdown is the storage. Obsidian is the editor. A Python CLI (memex) is the engine. No cloud. No lock-in. Everything the engine does is deterministic except where an LLM is explicitly invoked, and those calls are dry-run by default, snapshotted before apply, and cite their sources by slug.

• • •

The flow

  inbox/ → raw/ → wiki/ → projects/ → outputs (_qa, _essays, _slides, _index, _charts)     ↑                        ↑             ↑     └─────── feedback ───────┴─────────────┘

Capture into inbox/ — fast, unstructured, no decisions. Promote to raw/ during the weekly review — full text, citation metadata. Compile with memex compile — LLM reads raw, proposes atomic notes for wiki/. Link — every wiki note links to at least two others. Promote status: seed → draft → evergreen as notes earn trust. Produce on demand — memex qa, memex export essay, memex export slides. Feed back — every output surfaces gaps; you fill them; the vault compounds.

• • •

The folders

<your-vault>/  <your-name>.md       persona core — always in LLM context  memexlab.md        system core — always in LLM context  inbox/                 fast capture  raw/                   sources with citation metadata  wiki/                  atomic canonical notes — your thinking  people/ companies/ philosophies/ eras/   ontology entities  projects/<slug>/       goal-driven workspaces with end dates  archive/               what no longer earns active space  templates/             Templater frontmatter  _qa/ _index/ _essays/ _slides/ _charts/ _lint/   engine outputs  .memex/                snapshots, log, state

• • •

The engine

memex doctor    config sanity checkmemex migrate   bring frontmatter up to schema (dry-run + snapshot + apply)memex rollback  inverse of migrate --applymemex compile   raw source → atomic notes (LLM)memex lint      validate frontmatter, links, stubsmemex qa        answer a question with cited sources (LLM)memex index     reading order / concept map / problem view (LLM)memex export essay   topic → full essay (LLM)memex export slides  topic → Marp-compatible deck (LLM)memex chart     tag frequency / type distribution / latticework coverage / timeline

Universal contract. Dry-run by default; --apply to write; snapshots before any migration; one JSON event to stderr per invocation; exit 0 on success, 1 on error.

• • •

The schema

Three frontmatter types, dispatched by folder:

  • AtomicFrontmatter — wiki and engine outputs. Required: date, type, status, tags, latticework.

  • CuratedFrontmatter — ontology, raw sources, projects. Required: title, type, created, updated, status, tags, latticework.

  • CoreFrontmatter — two root files. Required: same as curated plus whitelisted type values (persona | homepage | manifest).

Atomic type values. article | qa | lint | index | essay | slides | chart. Statuses. seed | draft | evergreen (atomic); seed | draft | active | evergreen (curated). Latticework problems. problem-1 (seeing reality) through problem-5 (playing long games).

Unknown keys are rejected. Broken links are errors. Lint fails the run on any error finding.

• • •

The rhythm

Cadence

Effort

What

Daily

automated

scripts/lint_daily.py writes a dated report; fix errors same-day

Weekly

60–90 min

Inbox sweep → raw triage → compile queue → link pass → lint review

Monthly

2–3 hr

Merge duplicates, promote drafts, refresh ontology, coverage charts

Quarterly

4–6 hr

One essay on a major topic; feedback gaps into the vault

Yearly

1 day

Full ontology audit; tag taxonomy review; schema wish list

• • •

The non-negotiables

  • Dry-run before --apply, always.

  • Read every compile output before applying.

  • Every atomic note carries provenance: source: or cited_slugs:.

  • Status transitions are manual; no auto-promotion.

  • Links into archive/ from active notes are forbidden.

  • API key in shell environment only; never in any file under version control.

• • •

The domains this vault is tuned for

Programmable money · stablecoin settlement · AI-native banking · embedded finance · financial infrastructure · regulatory strategy (MiCA, CBDC frameworks, GCC) · · · Islamic finance compatibility.

The two core notes anchor every LLM call in these domains, which is why outputs sound like you, not like a generic PKM assistant.

• • •

The compounding claim

Month 1: you have a working capture habit and a handful of atomic notes. Month 3: the vault can answer most questions in your primary domains. Month 6: first-draft outputs from the vault are better than first-draft outputs from scratch. Year 1: the vault is infrastructure for your thinking — a research assistant that never sleeps, never forgets, and never leaves.

The system is boring most days. It compounds.

Docs index · Overview →