Skip to content

AGENTS.md

Latest commit

 

History

History
35 lines (30 loc) · 3.11 KB

AGENTS.md

File metadata and controls

35 lines (30 loc) · 3.11 KB

Learned User Preferences

Code quality & process

  • Fix root causes for all users; no case-by-case patches, retries, or workarounds; revert speculative changes promptly
  • All config via env vars with sensible defaults; never hardcode IPs or machine-specific values
  • Repo code, docs, and comments in English; extracted CEO Brain knowledge preserved in the user's original language
  • Batch commits by logical grouping; commit and push only when explicitly asked; release only via git tags + CI

Architecture & design

  • Plan and discuss architecture before coding; deliver plans in current chat mode — don't switch to Plan mode unless asked
  • All interfaces (CLI, MCP tools, plugins) must expose the same capabilities as the HTTP API; never access chDB directly
  • Use asyncio.to_thread() for all blocking calls (chDB, embedding, LLM) inside the async server
  • Prefer event-driven hooks over periodic cron; hooks source code in project tree (cursor-hooks/), not .cursor/
  • Distinguish project-scoped facts from global principles; use project_id + score boosting to prevent cross-project knowledge pollution
  • AGENTS.md bullets parsed directly as principles (not re-extracted through LLM); CLAUDE.md goes through LLM extraction

Workflow

  • Deploy changes AND verify end-to-end yourself; don't tell the user to verify
  • Test with representative cases before bulk operations; import data newest-first; don't proactively scan beyond what's specified
  • Coordinate parallel sessions: if another session implements a feature, revert speculative changes and wait for merge

Learned Workspace Facts

  • ClickMem: local-first CEO Brain for AI coding agents; shared by Claude Code, Cursor, OpenClaw via MCP + REST API
  • CEO Brain: projects, decisions (with reasoning), principles (with confidence + evidence), episodes (TTL 180d), raw_transcripts; dedup gate prevents duplicates on insert
  • memory setup one-click: install service + hooks + discover agents + import history (background); memory status tracks progress
  • memory import reads Claude Code JSONL + Cursor agent-transcripts + CLAUDE.md/AGENTS.md; newest-first; incremental via ~/.clickmem/import-state.json
  • Project-scoped recall: same-project results boosted 1.3x, global 1.0x, other-project penalized 0.6x; prevents cross-project knowledge pollution
  • chDB (embedded ClickHouse); ReplacingMergeTree(updated_at); single-process lock; single port 9527 (REST + MCP SSE + hooks)
  • Local LLM: Qwen3.5 4-bit on MLX (8GB→2B, 16GB→4B, 32GB→9B); enable_thinking=False for structured JSON; embedding Qwen3-Embedding-0.6B (256d, never MPS)
  • Chunked extraction: conversations segmented at turn boundaries (max 5 × 4000 chars); dedup merges across segments; startup dedup cleans existing principles
  • PyPI clickmem; CI on Python 3.10/3.12/3.13; release via v* tags + Trusted Publisher
  • Config: CLICKMEM_SERVER_HOST/PORT (9527), CLICKMEM_REMOTE, CLICKMEM_API_KEY, CLICKMEM_DB_PATH, CLICKMEM_LLM_MODE/MODEL/LOCAL_MODEL
  • Tests: MockEmbeddingEngine + MockLLMComplete via conftest; in-memory chDB
  • Deploy: Mac Mini M4 32GB (mini.local) via Tailscale; launchd service with auto-restart