Display Name
Project Brain
Category
Agent Skills
Sub-Category
General
Primary Link
https://github.com/OoneBreath/claude-code-project-brain
Author Name
Slawomir Luzny
Author Link
https://github.com/OoneBreath
License
MIT
Other License
No response
Description
Project Brain is a persistent, navigable memory for Claude Code: a .project-brain/ folder of plain markdown with a small index the agent reads first and topic files it opens on demand. It records each project's stack, decisions, status and version history (across many projects in one place), so the agent stops re-reading large docs and stops attributing one project's facts to another. It makes no network calls and has no telemetry; the helpers are stdlib Python only.
Validate Claims
Clone the repo and run ./install.sh (a short, commented Bash script that copies the skill into ~/.claude/skills/ and, with consent, merges an optional Stop-hook into ~/.claude/settings.json after printing a diff and backup). In any project run /project-brain init: it writes .project-brain/index.md from cheap signals (package.json/pyproject/git) WITHOUT reading source. Everything is local files; index.compact is a deterministic generated view (python3 brain-compact) and brain-check validates the brain and flags contradictory facts. No network calls are made.
Specific Task(s)
Demonstrate that the agent recalls prior work and refuses to silently redo finished work: after init, save one unit of work as a 'verified' topic, then in a later session ask the agent to redo that same task.
Specific Prompt(s)
Run /project-brain init in a repo. Then: 'Save what we just did to the brain as a verified topic.' In a fresh session ask: 'How did we solve X?' followed by 'Re-do X.' The agent should read the index, surface the existing verified topic with its version history, and ask before redoing it instead of starting over.
Additional Comments
Zero runtime dependencies for the brain (plain markdown; helpers are stdlib Python). MIT. No telemetry, no network calls, no --dangerously-skip-permissions. brain-export bundles the brain for pasting into other assistants and redacts infra (IPs/paths/secrets) by default. Same brain works across Claude Code, Cursor and Windsurf. Traction: 210 clones (83 unique cloners) in the last 14 days with no paid promotion.
Recommendation Checklist
Display Name
Project Brain
Category
Agent Skills
Sub-Category
General
Primary Link
https://github.com/OoneBreath/claude-code-project-brain
Author Name
Slawomir Luzny
Author Link
https://github.com/OoneBreath
License
MIT
Other License
No response
Description
Project Brain is a persistent, navigable memory for Claude Code: a .project-brain/ folder of plain markdown with a small index the agent reads first and topic files it opens on demand. It records each project's stack, decisions, status and version history (across many projects in one place), so the agent stops re-reading large docs and stops attributing one project's facts to another. It makes no network calls and has no telemetry; the helpers are stdlib Python only.
Validate Claims
Clone the repo and run ./install.sh (a short, commented Bash script that copies the skill into ~/.claude/skills/ and, with consent, merges an optional Stop-hook into ~/.claude/settings.json after printing a diff and backup). In any project run /project-brain init: it writes .project-brain/index.md from cheap signals (package.json/pyproject/git) WITHOUT reading source. Everything is local files; index.compact is a deterministic generated view (python3 brain-compact) and brain-check validates the brain and flags contradictory facts. No network calls are made.
Specific Task(s)
Demonstrate that the agent recalls prior work and refuses to silently redo finished work: after init, save one unit of work as a 'verified' topic, then in a later session ask the agent to redo that same task.
Specific Prompt(s)
Run /project-brain init in a repo. Then: 'Save what we just did to the brain as a verified topic.' In a fresh session ask: 'How did we solve X?' followed by 'Re-do X.' The agent should read the index, surface the existing verified topic with its version history, and ask before redoing it instead of starting over.
Additional Comments
Zero runtime dependencies for the brain (plain markdown; helpers are stdlib Python). MIT. No telemetry, no network calls, no --dangerously-skip-permissions. brain-export bundles the brain for pasting into other assistants and redacts infra (IPs/paths/secrets) by default. Same brain works across Claude Code, Cursor and Windsurf. Traction: 210 clones (83 unique cloners) in the last 14 days with no paid promotion.
Recommendation Checklist