project-memory（项目记忆）

project-memory

Slash Commands

/memory · /update-memory · /save · /handoff · /continue

Two Modes

Auto-detect: Fresh session (no prior turns) = Handoff. Mid-session = Memory. User can force either mode with explicit slash commands.

Red Flags

These thoughts mean STOP — you're skipping the skill:

Handoff Mode

Run on fresh sessions, agent switches, or "continue" / "handoff".

Step 1 — Scan all agents

# Use `python` instead of `python3` on Windows
python3 <skill-dir>/scripts/scan_agent_state.py "$(pwd)"

[READ] references/agent-parsing-gotchas.md — WHEN script fails, fall back to manual reads

Step 2 — Consolidate

[DO] Output one line per active agent:

[agent] — [last active] — [doing what] — [stopped because] — [next action]

Step 3 — Update memory files

[DO] Run the Update Algorithm below with findings from the scan.

Step 4 — Resume

[DO] Identify the most recently active agent's unfinished work [DO] If multiple agents worked on different things → ask user which to continue [DO] If the work needs a specific tool (e.g., Codex MCP) → say so, don't pretend [DO] Execute the next action from consolidated state

Memory Mode

Run mid-session, on /memory, /save, or every 5–7 turns as hygiene.

[DO] Read 3 structured files + git log → run Update Algorithm

Update Algorithm

[READ] references/initialization.md — WHEN files are missing or first run

1. [DO] Read DECISIONS.md, CONTEXT.md, STACK.md (from wherever they live)
2. [DO] Get ground truth via git: git log --oneline -10 git diff --stat HEAD~5 2>/dev/null || git diff --stat HEAD Use git as truth. If not a git repo or shallow clone: skip git, rely on agent scan + files.
3. [DO] If Handoff mode: also use scan results from Step 1
4. [DO] Extract:

- New decisions → append to DECISIONS.md - New failures/root causes → append to STACK.md do-nots - Version changes → update STACK.md locked versions - Current state → overwrite CONTEXT.md - Learned knowledge → write to memory/ as atomic files

[READ] references/decisions-format.md — WHEN writing to DECISIONS.md [READ] references/context-format.md — WHEN writing to CONTEXT.md [READ] references/stack-format.md — WHEN writing to STACK.md [READ] references/knowledge-graph.md — WHEN writing learned knowledge to memory/ [READ] references/token-budget.md — WHEN any file approaches its line cap

1. [DO] Write only changed files
2. [DO] Report: "Updated X, Y. Z unchanged."

Agent Config Hygiene (every invocation)

[READ] references/agent-config-hygiene.md — WHEN checking agent config structure [DO] Run silently on every invocation as final step:

• Detect agent config files at project root (AGENT.md, CLAUDE.md, GEMINI.md)
• If any is a dump (>40 lines) → extract knowledge to memory/, leave lean protocols
• Verify symlinks: CLAUDE.md and GEMINI.md should be relative symlinks to AGENT.md

Rules Before Writing

1. Read all three files first. Never update blind.
2. Update only what changed. One new decision → one new entry.
3. CONTEXT.md fully overwrites. The other two append only.
4. Failures are first-class. A root cause captured = a dead end avoided.
5. If nothing changed, say "No updates needed" and stop.

Edge Cases

[READ] references/edge-cases.md — WHEN encountering an unexpected scenario

References

• references/agent-parsing-gotchas.md — agent state format pitfalls
• references/agent-config-hygiene.md — symlink + dump detection + repair
• references/knowledge-graph.md — memory/ structure and conventions
• references/initialization.md — file creation protocol
• references/decisions-format.md — ADR format + template
• references/context-format.md — CONTEXT.md format + template
• references/stack-format.md — STACK.md format + template
• references/token-budget.md — line caps + archive protocol
• references/edge-cases.md — edge case table
• scripts/scan_agent_state.py — automated agent state scanner