Second Brain: A Local Knowledge Base That Reads Your Docs So You Don't Have To

Ingest once, retrieve automatically

From raw files to heading-aware chunks

Second Brain reads across roughly 19 file extensions - Markdown, TXT, PDF, DOCX, EPUB, and other common formats. Google Drive files are exported to Office formats before ingestion so they go through the same pipeline as everything else.

As of 2.3.2, chunking is heading-aware: instead of slicing documents into arbitrary fixed-length blocks, it splits along document structure, which keeps related content together and makes individual chunks more meaningful as retrieval units.

You can drive ingestion directly with the CLI, or let a hook do it per-prompt: a retrieval step pulls the top relevant excerpts and injects them as `[SECOND_BRAIN]` context before the model responds. The always-on dashboard keeps that retrieval path warm, with lookups landing around 60ms.

# Ingest a document into the project's knowledge base
monomind doc ingest ./docs/architecture.pdf

# Search it directly from the CLI
monomind doc search --query "how does retrieval merge project and global results"

# List everything currently indexed
monomind doc list

One brain per project, one brain for everything

Project and global knowledge, merged at query time

Every project gets its own local index, but monomind also maintains a global, cross-project brain at `~/.monomind/global-brain` (the location is configurable via `MONOMIND_GLOBAL_BRAIN_DIR`). It's deliberately kept as a sibling of the per-project store, so it's never accidentally swept up by `cleanup --data`.

Ingesting a path outside the current project routes to the global brain automatically - there's no separate command to remember. At query time, retrieval merges results from both stores: when a project result and a global result tie, the project result wins, and global hits are labeled `[global]` so you always know where an excerpt came from.

For programmatic access, the MCP server exposes `knowledge_search` (which accepts a `store: "project" | "global"` filter when you want to search just one), plus `knowledge_ingest` and `knowledge_remove` for adding and retracting documents. `knowledge_remove` hides a document from search immediately and is reversible by re-ingesting it. Telemetry about queries - never prompt text - is logged locally to `.monomind/metrics/second-brain.jsonl`.

“Retrieval merges project and global stores automatically - project wins ties, global hits are labeled [global] - so you never have to remember which brain you're searching.”

- Monomind Core Team

Measured, not just marketed

A CI-enforced recall bar and a single outbound call

It's easy to ship a search feature and never check whether it actually finds the right thing. monomind maintains a paraphrase golden-set eval - a fixed set of query/document pairs phrased differently than the source text - and enforces an 80% recall bar in CI. If retrieval quality regresses below that bar, the build fails.

On the network side, Second Brain's only outbound call is a one-time fetch of the embedding model (roughly 90MB) from HuggingFace the first time you index a document. After that, indexing and retrieval run locally.

The full CLI surface is under `monomind doc`, with subcommands for ingest, search, list, export, remove, reconcile, and eval - so you can inspect, audit, and re-run quality checks against your own indexed documents at any time.

{
  "tool": "knowledge_search",
  "arguments": {
    "query": "how is the global brain kept separate from cleanup --data",
    "store": "global"
  }
}