One Brain.
Any Tool.

Fair — if you live in one tool and keep disciplined rules files, you already have a lot of this. Three things a specs/folder doesn't do:

1. Nobody has to write the files.The git hook and file watcher capture what actually happened — decisions, commits, what you tried and abandoned. Hand-maintained docs drift the moment you get busy. Auto-captured memory doesn't.
2. Retrieval, not injection.A docs folder grows forever and you can't paste all of it into context. Cerebrumma embeds every memory and surfaces only what's semantically relevant to the task at hand — which matters exactly when the project gets big enough that “include everything” stops working.
3. It distills.The dream cycle compresses raw episodic logs into lasting rules and facts, promoting them into deeper memory layers. That's the part nobody does by hand.

Yes — and they will, if you point them there. The problem isn't reading, it's knowing which slice matters right now, and keeping it current. Every tool also has its own convention (.cursor/rules, CLAUDE.md, …) and none of them share. Cerebrumma serves the relevant memory over MCP, so it's the same Brain whichever tool asks.

No. Cerebrumma runs entirely locally by default. Embeddings are computed on-device, and every memory is a plain Markdown file inside your project. No keys, no cloud, no telemetry. The only time anything leaves your machine is if you opt in to automatic reflection (dream --auto) — and you supply that key.

No. Everything core — capture, search, the neural map, the manual dream cycle — works fully offline. A key only enables the automatic dream cycle. Without one, you run the reflection through any LLM you already pay for.

No. Auto-reflection is frequency-capped by design: it only runs once enough new episodic memory has accumulated, and it enforces a cooldown between runs. It can't spam your provider in the background.

No. Memories are small plain-Markdown files with YAML frontmatter. The embedding model is cached once, globally, in ~/.cerebrum/models — never inside your project.

It reads your raw episodic memory (commits, sessions, saves) and distills it into lasting knowledge — facts, skills, and preferences — promoting each into the right memory layer. Run it by hand with any LLM, or let the daemon do it in the background. It never deletes anything automatically; prune suggestions are always left for you to review.

Anything that speaks MCP — Claude Code, Cursor, Antigravity, Windsurf, and more — plus the cerebrumCLI itself. Your Brain isn't tied to any one of them.

Those store notes you write and link by hand. Cerebrumma captures what you actually did, automatically, and serves it to your AI. There's no manual linking — connections are inferred from meaning. It complements a notes app; it doesn't replace one.