/install autobiographical-memory
Autobiographical Memory
Core Concepts
Memory has two complementary layers:
| Layer | File | What it stores |
|---|---|---|
| Episodic | memory/YYYY-MM-DD.md |
Raw daily events, conversations, decisions, observations |
| Semantic | MEMORY.md |
Curated knowledge: user preferences, facts, lessons, identity |
The memory lifecycle: Capture → Consolidate → Recall → Review
Quick Start
# Recording an event (episodic)
Append to `memory/YYYY-MM-DD.md`:
- Met with [person] about [topic]. Decision: [outcome].
- User prefers [preference]. Updated MEMORY.md.
# Recalling before responding
1. Run `memory_search` with relevant keywords
2. If results are thin, read recent `memory/YYYY-MM-DD.md` files
3. Check `MEMORY.md` for long-term facts
# Consolidating (periodic maintenance)
1. Read recent daily files (last 7-30 days)
2. Extract significant items → update MEMORY.md
3. Remove stale entries from MEMORY.md
Episodic Memory — Daily Notes
What to Record
Always write to daily notes for:
- Decisions with rationale: "Chose X over Y because Z"
- User preferences discovered implicitly or explicitly
- Important conversations — summary, not transcript
- Mistakes & lessons — what went wrong, what to do differently
- Project milestones — what was done, what's blocked
- Identity changes — if SOUL.md, USER.md, or other self-files changed
What to Skip
- Routine operations ("checked email, nothing new")
- Transient states ("feeling tired")
- Content better stored elsewhere (code snippets in projects, API keys in config)
- Something the user explicitly said doesn't matter
Format Convention
## Events
- [event description]
## Decisions
- [decision + rationale]
## Observations
- [insights or patterns noticed]
## Notes
- [anything else worth remembering]
Semantic Memory — MEMORY.md
Structure
## User Preferences
- Directly stated preferences without inference
## Project Context
- Active projects and their status
## Relationships & People
- Key people, roles, context
## Technical Environment
- Tools, config, quirks discovered
## Lessons Learned
- Mistakes to avoid, patterns that work
When to Update MEMORY.md
- User states a clear preference
- A project direction is set
- A mistake teaches a lesson worth preserving
- Every few days during heartbeat: consolidate from daily notes
When to Remove from MEMORY.md
- Project is done / abandoned
- Preference was explicitly reversed
- Information is now obvious context (e.g. "user speaks Chinese" — that's already in USER.md)
- Stale for >3 months without reference
Consolidation Workflow
Suitable for heartbeat routines. Do this every 3-7 days:
1. List memory/*.md, sort by date (newest first)
2. Read files since last consolidation
3. For each significant item:
a. Is it already in MEMORY.md? → Skip or update
b. Is it transient? → Skip (leave in daily note)
c. Is it important? → Add to MEMORY.md
4. Read MEMORY.md for stale entries → remove or archive
5. Write updated MEMORY.md
Recall Strategy
Before answering questions about prior work, people, preferences, or context:
- Search first:
memory_search(query="relevant terms")— this searches both daily notes and MEMORY.md - Narrow scope: If search returns weak results, try multiple query phrasings
- Deep dive: For specific periods,
memory_get(path="memory/YYYY-MM-DD.md")to read raw daily notes - Cross-reference: Check USER.md, SOUL.md, TOOLS.md for identity/preference info
When Recall Fails
- Say clearly "I checked my records and don't have information about that"
- Don't fabricate memories
- If the user says "don't you remember? I told you X" — apologize and record it properly this time
The Forgetting Curve
Not everything needs to persist. Use these filters:
- Keep in daily notes: Everything noteworthy for 30-90 days
- Promote to MEMORY.md: Only what's likely to be needed again
- Delete/archive: What's clearly obsolete after review
References
- For detailed consolidation scripts: see references/consolidation.md
- For memory search patterns: see references/recall-patterns.md
Scripts
scripts/consolidate.py— Scan recent daily notes and suggest MEMORY.md updatesscripts/stats.py— Memory file statistics (sizes, dates, coverage)
- 确保已安装 OpenClaw(本地或 Docker 部署)
- 在对话框中输入安装命令:
/install autobiographical-memory - 安装完成后,直接呼叫该 Skill 的名称或使用
/autobiographical-memory触发 - 根据 Skill 的参数说明提供必要输入,即可获得结构化输出
Autobiographical Memory 是什么?
Structured personal memory system that enables agents to persist, consolidate, and recall episodic and semantic memories across sessions. Use when: (1) recor... 它是一个面向 Claude Code / OpenClaw 的 AI Agent Skill 插件,目前累计下载 25 次。
如何安装 Autobiographical Memory?
在 OpenClaw 或 Claude Code 对话框中运行命令「/install autobiographical-memory」即可一键安装,无需额外配置。
Autobiographical Memory 是免费的吗?
是的,Autobiographical Memory 完全免费,采用 MIT-0 许可证,可自由下载、安装和使用。
Autobiographical Memory 支持哪些平台?
Autobiographical Memory 跨平台运行,可在任意部署了 OpenClaw / Claude Code 的环境中使用(cross-platform)。
谁开发了 Autobiographical Memory?
由 xuyuchen(@xuyucheneureka)开发并维护,当前版本 v1.0.0。