Semantic search (srag)

brain_recall is keyword/tag search — fast, deterministic, but you have to know what you’re looking for. For “what did we figure out about X” questions where the right note’s title doesn’t contain X, semantic search wins.

The plugin ships an integration with srag — a local Rust+Python tool that does chunking + sentence-transformer embeddings + on-disk vector search. Brain notes are exported as markdown, srag indexes them as a synthetic claude-brain project, and brain_search_semantic queries the index.

Why this layer

Tool	Best for	How it matches
brain_recall	”Find the note titled X” / known tag	substring + Trilium label query
brain_search_semantic	”What did we decide about idempotent migrations”	sentence embeddings, cross-section recall
srag (direct)	“Where in the codebase is the JWT secret rotated”	embedded code chunks across many repos

The three layer up nicely: keyword for known notes, semantic for vague recall, srag direct for code.

Setup

# 1. Install srag (one-time, per host)
curl -fsSL https://raw.githubusercontent.com/wrxck/system-rag/main/install.sh | bash

# 2. Build/install the brain MCP (already done if you used the plugin install wizard)

# 3. From inside Claude:
brain_sync_srag()

brain_sync_srag does:

1. Walks the brain via the active adapter’s listAll()
2. Writes each note to ~/.local/share/claude-brain/srag-export/<backend>/<sanitised-id>__<title>.md with YAML frontmatter (id / title / path / modifiedAt)
3. Calls srag index --name claude-brain <export-dir> to chunk + embed

Re-run after batches of new notes; it’s idempotent.

Tools

brain_sync_srag

Parameter	Type	Default	Description
clean	bool	false	Wipe the export dir first (full rebuild). Otherwise incremental.
underPath	string	(all)	Restrict to notes under this path — e.g. "Claude Memory/Standards".

Returns a summary: notes written / skipped / errors, plus the srag index output.

brain_search_semantic

Parameter	Type	Default	Description
query	string	—	Natural-language question.
limit	number	8	Max snippets returned.

Returns ranked snippets with noteId (parsed from the export filename) so Claude can chase a hit back to the canonical note via brain_recall or the backend’s UI.

Discipline pattern

The recommended recall flow when the user asks an open question:

1. brain_search_semantic first — broad cast.
2. For any hit, follow up with brain_recall by id/title to load the full canonical note.
3. Cite the result by backend / category / title / last-modified date before acting.

The plugin’s _discipline skill enforces this pattern automatically.

Limitations

• Embedding cost: srag uses sentence-transformers locally (CPU). First index of ~500 notes takes a minute or two; queries are sub-second.
• Adapter scope: listAll() returns all notes tagged #claude-brain for Trilium, all .md files under the vault for Obsidian, and the integration’s accessible pages for Notion (Notion is rate-limited; capped at 200).
• Stale until you sync: there is no live watcher — re-run brain_sync_srag (or wire it to a cron / systemd timer) after meaningful note batches.

Env

Variable	Default	Description
BRAIN_SRAG_EXPORT_DIR	~/.local/share/claude-brain/srag-export	Where the markdown export lives.
SRAG_BIN	srag (PATH)	Override the srag binary path.