memory

name	memory
description	Layered workspace memory — facts, structured event/focus, and grepable history.
always	true

Memory

Structure

File	Role	In system prompt
memory/MEMORY.md	Long-term facts (preferences, project context, relationships).	Yes — always loaded when non-empty.
memory/EVENT_MEMORY.md	Structured event / focus memory (显性/隐性关注点、时间线、演化).	Yes — always loaded when non-empty (same pipeline as MEMORY.md).
memory/HISTORY.md	Append-only event/decision log.	No — search on demand (grep / exec). Each entry should start with [YYYY-MM-DD HH:MM].

• Bootstrap EVENT_MEMORY: copy the template from templates/memory/EVENT_MEMORY.md into memory/EVENT_MEMORY.md when the user (or you) wants structured tracking of what the user is focused on across sessions — not only raw facts in MEMORY.md.

EVENT_MEMORY.md (when used)

Follow the template sections:

1. Current focus — Explicit vs implicit user goals; update only when aligned with the current thread (关注ID CF-*).
2. Historical focus log — Append-only; recent vs archived early focuses; shorten old items over time; merge many small focuses into one line when the list grows.
3. Focus evolution tracker — Short changelog: 新增 / 深化 / 转移 / 消退 / 撤销 (use the diff-style block in the template).

• compress_session tool: summarizes the unconsolidated chat tail and replaces it in the session JSONL only — it does not write MEMORY.md or HISTORY.md.
• Token-based archival (when the prompt exceeds the context window): still archives chunks into MEMORY.md and HISTORY.md via the internal save_memory pipeline. It does not fill EVENT_MEMORY.md. You maintain EVENT_MEMORY.md with edit_file / write_file when the skill applies.

Search past content

HISTORY.md is not injected into the prompt; use search when it grows.

Choose the method by file size:

• Small files: read_file, then filter in memory.
• Large or long-lived HISTORY.md: use exec for targeted search.

Examples:

• Linux/macOS: grep -i "keyword" memory/HISTORY.md
• Windows: findstr /i "keyword" memory\HISTORY.md
• Cross-platform: python -c "from pathlib import Path; p=Path('memory/HISTORY.md'); ..."

Prefer targeted CLI search for large logs. For EVENT_MEMORY.md, the model usually already sees it in context; use grep only if you must search a copy or an archived export.

When to update what

MEMORY.md — Important durable facts as soon as they are stable:

• User preferences (“I prefer dark mode”)
• Project facts (“The API uses OAuth2”)
• Stable relationships (“Alice is the project lead”)
• Valuable execution experience

HISTORY.md — Session archival and grep-friendly narrative (often filled by consolidation). You normally do not hand-edit unless adding a manual note.

EVENT_MEMORY.md — When you need a focus-centric view (topic shifts, multi-session threads, explicit vs inferred intent):

• After a topic pivot, consider updating current focus and appending to the evolution tracker.
• When early focuses pile up, summarize per template (近期 → 早期/归档).
• Keep §2 append-only except for explicit conflicts or staleness, as described in the template.

Auto-consolidation

When the session grows large, token-based archival writes to HISTORY.md and MEMORY.md. The compress_session tool only trims the session log; use memory_update or edit_file if you want those files updated manually. You do not need to duplicate archival into EVENT_MEMORY.md unless you are maintaining the structured focus model.