This is the solution to Claude forgetting everything between sessions and burning thousands of tokens on context re-explanations. It's a three-layer memory system (global, project, branch-aware) that compresses session state into under 2,000 tokens, verified to save 67% on warm-up cost. When you start a session, it auto-loads, diffs against your git history, flags stale references, and greets you with exactly where you left off. Mid-session it updates on "remember this" or milestones. On wrap-up it compresses completed work and archives what's done. The practical win is sessions that last 3-5x longer because you're not re-explaining the same architecture decisions every time. Activates automatically if MEMORY.md exists, or on any variation of "pick up where we left off."
npx -y skills add nagendhra-web/memory-bank --skill memory-bank --agent claude-codeInstalls into .claude/skills of the current project.
Select a file.
juliusbrussee/caveman
mattpocock/skills
shadcn/improve
obra/superpowers
forrestchang/andrej-karpathy-skills
vercel-labs/skills