MIT · local · markdown

Shared memory for
local coding agents.

One wiki. Every harness on the machine consults it and lightly updates it while doing real work. A named brain-keeper skill runs a batch “dream” only when you ask.

A dream cycle reads every local transcript it can find. Grok, Claude Code, Codex, OpenCode, Pi, Warp, Cursor, Gemini. Those files can contain API keys, customer names, and private data. The standing rule is cite-or-omit: durable facts go in the wiki, secrets do not.

01Always-on versus dream

Daily agents keep a real task. The wiki is a side effect. They read matching notes and merge durable facts. They do not run a full distillation unless you say so.

Always-on rule

Installed into each detected harness as brain.md or a marked block in existing AGENTS.md / WARP.md / CLAUDE.md. Those files are appended, never replaced.

Dream cycle

Say Update the brain. in any wired harness. That invokes brain-keeper, crawls closed sessions newer than the last changelog date, and writes only under the wiki folder you chose.

02Install

Needs bash and python3. Optional: gum for nicer prompts. The nightly timer needs systemd (Linux). macOS still gets the wiki and harness wiring; run the dream by hand.

install.sh
$ git clone <source>
$ cd quickbrain
$ ./install.sh
  1. Creates a wiki folder (default ~/Quickbrain, or any path you type).
  2. Seeds How Quickbrain Works plus empty Projects, Concepts, Entities, Preferences, Learnings.
  3. Writes onboarding notes only for answers you actually gave. Every prompt is skippable.
  4. Detects coding harnesses already on this machine and installs the rule + skill.
  5. Leaves the nightly timer off unless you pass --nightly.

Non-interactive:

flags
./install.sh --yes --wiki ~/Notes/brain \
  --name "Ada" --work "independent developer" \
  --projects "personal site, invoice tool" \
  --people "Nico" \
  --prefs "Prefer small diffs and cite sources"

Re-run after you install a new harness. Uninstall removes wiring and keeps the wiki: ./uninstall.sh

03The wiki is yours

Plain markdown. Point install at a new folder, an Obsidian vault, a git repo, iCloud, Syncthing — whatever you already sync. Config lives at ~/.config/quickbrain/config. Relocate the folder, update that path, re-run ./install.sh.

This installer does not ship anyone’s personal notes. The wiki it creates on your machine is empty plus the how-it-works page, then whatever you and your agents write.

04Harnesses

Detected by the harness config directory, or by a uniquely named CLI on PATH. Pi is directory-only (~/.pi/agent) because pi is a common command name. Missing or cloud-only transcript stores are skipped during a dream, not invented.

HarnessClosed conversations
Grok Build~/.grok/sessions/
Claude Code~/.claude/projects/
Codex~/.codex/sessions/ and state_5.sqlite
OpenCode~/.local/share/opencode/opencode.db
Pi~/.pi/agent/sessions/
WarpUsually cloud; local files if present
CursorLocal transcripts under ~/.cursor/ if present
Gemini CLILocal logs under ~/.gemini/ if present

05Nightly timer is optional

Off by default. ./install.sh --nightly enables a Linux systemd user timer around 03:07 that runs grok -p (or claude -p) with the dream prompt. Only do this on a machine you control. There is no launchd installer. On macOS say Update the brain.

06What this is not

Not a cloud brain

Nothing is uploaded to us. Transcripts stay on the machine. Cloudflare only serves this static site.

Not a second PKM product

The distinctive piece is coding-agent continuity: work as source, always-on versus dream, harness adapters, cite-or-omit. Obsidian is optional.

Day to day: work as usual. To distill closed sessions, say Update the brain.