LLM Brain

The brain your AI maintains.

Eva turns curated sources into a maintained, interlinked body of knowledge on your Mac. Local files, local Git, your agent.

Download for macOS View on GitHub

Retrieval re‑reads. Eva remembers.

A retrieval pipeline answers every question from scratch: chunk the documents, fetch what looks relevant, synthesize, discard. The work of understanding is thrown away the moment the answer is delivered.

Eva's agent writes what it learns into Markdown pages that link to each other: entities, concepts, source summaries, saved analyses. Every ingest and every answer worth keeping improves the same artifact, so the next question starts further ahead instead of starting over.

The brain is a plain folder. No vector database, no embeddings to rebuild. You can open it in Obsidian or any text editor, and it stays readable without Eva.

See it in action.

A real 117-page brain. A section cluster expands into its pages, then selecting a source summary fans red links to every concept it supports while the rest of the map dims.

Eva's graph view navigating a real brain: a cluster expands into pages, then a summary is selected, highlighting its linked concepts in red while everything else fades

Four operations, one review gate.

Ingest

Pick a source. Eva copies it into raw/, commits it, and the agent reads the existing wiki before writing: merging into pages that already exist, linking new ones from the index, citing the source. All of it happens on a Git branch in an isolated worktree.

held: 1 deletion deletions never auto‑merge

Query

Ask the maintained brain, not a pile of chunks. Answers cite the exact pages and raw sources that support them. Saving a useful answer proposes a reviewable analysis page; asking alone never changes anything.

concepts/compounding · raw/1977-letter.txt

Health

Deterministic lint catches broken links, orphan pages, and missing provenance. An optional agent pass, strictly read‑only, flags contradictions, stale claims, and coverage gaps. It reports; it never edits.

broken-link: [[lint-rules]] does not resolve

Tools

Turn the open brain into flashcards, a practice exam, an evidence map, a decision brief. Every tool is read‑only and source‑cited, and saving a result goes through the same review gate as everything else.

flashcards evidence map practice exam

Local by design.

A folder, not a service: Brains are Markdown folders in ~/Documents/Eva/Brains. No account, no cloud database, nothing to sign in to.
Bring your own agent: Eva drives the Codex CLI or Claude Code already signed in on your Mac. It never stores API keys or credentials.
Git is the review gate: Every brain is its own local repository. Agent work lands on a branch, and deletions or new structural issues always wait for your explicit accept or reject.
raw/ is immutable: Agents may read your original sources but never edit, move, or delete them. Eva verifies this on the working tree and on the committed diff before anything can merge.
Nothing leaves your Mac: Eva never creates a remote and never pushes. Backing up or sharing a brain is a choice you make outside Eva.
V1, macOS only: Built and tested on macOS alone; other platforms are a future decision, not a checkbox. Today you build it from source with Node.js, Rust, and a signed‑in agent CLI.