Agent Memory — Multi-Agent Coordination Protocol

A memory/ directory at .claude/memory/ (or ~/.claude/projects/<project>/memory/) gives multiple agents and sequential sessions a shared source of truth. Any agent entering a project reads from memory to get context, claims work, tracks tasks, and leaves a clean handoff.

File Structure

memory/
  MEMORY.md              # Index — brief, always loaded into context
  tasks.json             # Task tracker SSOT — all open/in-progress/done tasks
  active-work.md         # Live coordination — what's actively being worked on
  completed-work.md      # Archive of finished work with details
  worklog.md             # Rolling session history (last 30 sessions)
  decisions.md           # Architecture decisions with rationale
  gotchas.md             # Known bugs, workarounds, things to avoid
  projects/              # Per-project or per-topic state files