memex-md

You like Obsidian? Your LLM will love it too.

Memex: Vannevar Bush's 1945 concept of a "memory extender" - a device for storing and retrieving personal knowledge. The conceptual ancestor of personal wikis and second brains.

Fuzzy find, wikilink graph traversal, and semantic search for markdown vaults. Point it at your Obsidian vault (or any markdown folder).

Quick Start

uv tool install memex-md                                                    # find, explore, rename
uv tool install memex-md --with "memex-md[semantic]" --torch-backend=cpu    # + semantic search (CPU, ~200MB)
uv tool install memex-md --with "memex-md[semantic]" --torch-backend=auto   # + semantic search (auto-detect GPU)

If installed, you can run memex directly, or via the alias mx:

mx vault:add personal ~/notes ~/journal
mx search "How does the auth flow handle token refresh?" -v personal
mx find knn
mx explore auth personal

For a memex skill, see my agent workflows.

How It Works

A vault is a named collection of directories. Each vault has its own embedding model and SQLite index (~/.local/share/memex-md/<vault-name>/index.db).

The index contains:

• Wikilink graph for backlink/outlink queries
• Extracted frontmatter (aliases, tags)
• Embeddings for semantic similarity (optional, default: embeddinggemma-300m)

Indexing is incremental — on each command, only files with changed mtimes are re-indexed. Hidden directories (.obsidian, .trash, .git, etc.) are excluded.

Semantic search requires the semantic extra (pip install memex-md[semantic]). Without it, find, explore (outlinks/backlinks), and rename work normally — only search and the "similar notes" section in explore are unavailable.

Note: Initial embedding computation is GPU-intensive. Example: ~3800 notes took ~7 minutes on an RTX 3070 Ti.

Configuration

Vaults are configured via CLI. Config lives at ~/.config/memex/config.toml.

mx vault:add personal ~/notes ~/journal
mx vault:add work ~/work-docs --model some/other-model
mx vault:list
mx vault:info personal
mx vault:remove personal --path ~/journal   # remove one path
mx vault:remove work                        # remove entire vault

Re-add a vault with --model to change its embedding model (triggers re-embedding). Use --model none to disable semantic search (wikilink navigation only).

Embeddings run on a GPU when torch finds one. To pin every vault to one device, set it under [defaults]:

[defaults]
device = "cpu"   # any torch device: "cpu", "cuda", "cuda:1", "mps"

Commands

find

mx find knn
mx find "neural ordinary" -v personal
mx find auth -n 20

Fuzzy match against note titles, frontmatter aliases, and paths. No embeddings needed — instant results. Ranked: exact title/alias > substring > fuzzy (rapidfuzz WRatio). Multi-word queries match any part; more parts hitting = ranked higher.

search

mx search "How does the auth flow handle token refresh?" -v personal
mx search "What approaches did we consider for caching?" -v work --full

Embeds the query and ranks indexed notes by cosine distance. Natural language questions of a few sentences tend to work well.

explore

mx explore auth personal
mx explore docs/api-design work --full

Shows a note's outlinks ([[wikilinks]]), backlinks, and semantically similar notes.

note_path can be a title (auth) or path (docs/auth.md). Titles must be unique in the vault.

rename

mx rename old-name new-name personal
mx rename docs/guide manual work

Renames a note file and updates all [[wikilinks]] pointing to it. Handles path links, title links, aliases, and heading refs. Ambiguous links (multiple files share a name) are skipped with warning.

index

mx index
mx index -v personal

Trigger indexing. Runs automatically before search/explore.

Development

uv sync
make check          # ruff + ty
make test           # pytest
make release-patch  # 1.0.0 -> 1.0.1, tag, push
make release-minor  # 1.0.0 -> 1.1.0
make release-major  # 1.0.0 -> 2.0.0