Recoleta
GitHub
Source note

Project Brain – Persistent memory index for AI coding

Authors
Slav_fixflex
Source
Hacker News
Published
Collection
Software Intelligence
AI CodingProject MemoryClaude CodeCode IntelligenceDeveloper Tools
Open GitHubMarkdown source

Summary

Project Brain is a Claude Code skill that keeps project memory in a small Markdown index with linked detail files. It targets repeated context setup, project mix-ups, and wasted tokens in AI coding sessions.

Problem

  • Claude Code users often re-explain architecture, deployment, stack, history, and pitfalls at the start of each session.
  • Long always-loaded notes or READMEs can waste context, including cases where users paste a 1000-line README for routine tasks.
  • Multi-project work can cause the model to confuse stacks, prior decisions, and completed work across repos.

Approach

  • The skill creates a .project-brain/ folder with index.md as a small map and per-project topic files under projects/<project>/<topic>.md.
  • Claude reads the index first, then opens one linked topic file only when the user asks about that area.
  • init detects projects from signals such as git, package.json, and pyproject.toml, then adds a pointer in CLAUDE.md so future sessions load the map.
  • Status values such as verified, failed, and in-progress record outcomes, while superseded notes keep prior decisions instead of overwriting them.
  • A brain-check validator catches broken pointers, malformed frontmatter, and index-to-topic status drift.

Results

  • The excerpt reports no benchmark, ablation, user study, or measured token reduction.
  • Claimed token benefit: only index.md loads eagerly, so per-session context is bounded by the index size rather than the full history or a 1000-line README.
  • Claimed recall benefit: after 3 months away from a project, Claude can use stored stack, deployment, decisions, failed attempts, and verified work without the user pasting context again.
  • Claimed navigation behavior: a question such as “how did we solve the cache issue?” should load 1 topic file through 1 index pointer instead of reading the whole knowledge base.
  • Setup claim: run install.sh once per machine, run init once per workspace, and one brain can catalog many projects on a server.

Link