Maintain persistent project memory for AI coding agents

Maintains a persistent, plain-Markdown project memory bank in .lore/ at the project root, capturing the context that normally lives only in the original developer's head: architecture (ARCH - what the project is and how it's shaped), decisions (DEC - why a choice was made over alternatives, including the tradeoffs), and conventions (CONV - how code should be written and what to avoid). Seven commands cover the full lifecycle: init for one-time setup, detecting scope boundaries from workspace config across pnpm, Yarn, npm, Lerna, Nx, Rush, Cargo, Go, and Bazel monorepos, falling back to a single _global/ scope for non-monorepo projects; sync to record a change after a non-trivial feature, refactor, or bug fix, and to create new scope directories when a change introduces one, not init, which is reserved for first-time setup or an explicit start-over; query to answer "what's the convention" or "why was X chosen" from memory; audit to check whether memory has drifted from the actual code; compress to rebuild the top-level SUMMARY.md digest once entries exceed 500 or the last compression is over 30 days old; mirror to regenerate platform-specific files (CLAUDE.md, .cursorrules, and similar) from the canonical .lore/ content; and history to show the git commits behind a specific memory entry.

Load it when the user explicitly invokes lore, names one of its seven subcommands, references .lore/, or asks to record, recall, audit, sync, or compress project memory about decisions, architecture, conventions, or monorepo scopes - not on a bare "init," "compress," "audit," or "query" alone, since those phrases equally plausibly map to the coding agent's own native commands (Claude Code's /init//compact) or an unrelated task like a security audit or a SQL query. Once loaded for a session, some commands propose themselves against internal thresholds rather than requiring the user to name them explicitly - sync proposes when 50+ changed lines span 2+ directories, a top-level module or dependency was added or removed, or a new convention was just discussed in chat; compress appends a notice to a sync proposal once entries exceed 500, SUMMARY.md is missing, or the last compression was over 30 days ago; sync also emits an alert when an active entry conflicts with current code or a candidate change - but every proposal still requires explicit user acceptance, since the skill never mutates files silently. init, query, and history are always explicit and never self-propose.

Input is the codebase itself for init/sync/audit, or a natural-language question for query/history; output is Markdown files under .lore/: a top-level SUMMARY.md digest that new agents read first, _global/ or per-scope scopes/<name>/ ARCHITECTURE.md/DECISIONS.md/CONVENTIONS.md files, a draft/ directory holding init proposals pending confirmation, and an audit/ directory holding reports that never mutate the main files. mirror additionally regenerates platform-specific memory files (CLAUDE.md, AGENTS.md, .cursorrules) from the same canonical .lore/ source, with an .archive/ backup on each mirror rewrite, and can be set to auto-regenerate during compress via auto_mirror: true in .lore/.config.json. The steady-state cadence: init once, then sync/query/audit recurring, compress when the summary goes stale, mirror to publish structural changes.

Framework- and platform-agnostic: readable by any coding agent that can read files, and explicitly listed as compatible with Claude, Cursor, Gemini, Codex, Copilot, OpenCode, Cline, and Aider, mirroring its canonical content out to each platform's own memory-file convention. It ships helper scripts (id_hash, list_entries, find_duplicates, find_stale, history, documented in both English and Chinese) and a versioned .config.json#schema_version with its own migration and deprecation workflow for compatibility across lore releases.

Development teams and AI coding agents that need durable, cross-session project context - the why behind past decisions and the conventions a codebase actually follows - stored as plain Markdown any agent or human can read, rather than re-deriving architecture and rationale from scratch or scattering it across ephemeral chat history.