Basic Memory#

Purpose#

OpenClaw already stores memory in ~/.openclaw/workspace/MEMORY.md and loads it automatically at session start. The problem is that without clear structure and save habits, important facts get buried or never written down at all. This skill teaches the agent exactly what to save, where to put it, and how to retrieve it — so memory actually works in practice.

Storage Layout#

All readable memory lives under the workspace:

text

~/.openclaw/workspace/
├── MEMORY.md                  ← long-term facts, loaded every session
└── memory/
    └── YYYY-MM-DD.md          ← daily log, today + yesterday auto-loaded

Do not write to ~/.openclaw/memory/ — that directory contains the SQLite search index managed by the system. Only the workspace files are yours to read and write.

MEMORY.md Structure#

Organize MEMORY.md into four sections. Create any missing section on first write.

markdown

## Facts
- [2026-04-09] User's name is Billy. Works as an AI engineer at a startup.

## Preferences
- [2026-04-09] Prefers concise replies. Dislikes unnecessary preamble.

## Action Items
- [ ] [2026-04-09] Follow up on API integration by Friday.
- [x] [2026-04-08] Send draft proposal to team. (done)

## Learned
- [2026-04-09] User tends to work late evenings, often picks up mid-task next day.

Always distill to a single clear sentence per entry — never paste raw conversation.

On Session Start#

OpenClaw auto-loads MEMORY.md plus today's and yesterday's daily files. Use this loaded context immediately — no tool call needed. Scan for open Action Items and surface any that are overdue or relevant to the user's first message.

Auto-Save During Conversation#

Detect and save without being asked when the user shares:

SignalSectionFileName, job, company, location, family## Facts``MEMORY.md"I prefer…", "I hate…", "always use…"## Preferences``MEMORY.md"I decided to…", "let's go with…", deadlines## Action Items``MEMORY.mdProject names, key people, recurring contextdaily logmemory/YYYY-MM-DD.mdPatterns you've noticed about the user## Learned``MEMORY.md

After auto-saving, confirm briefly in one line: "Noted: [distilled fact]."

Retrieving Memory#

Use the built-in tools when searching:

• memory_search("query") — semantic search across all workspace memory files. Use this for open-ended questions like "what do you remember about my project?"
• memory_get("MEMORY.md") — read the full long-term memory file directly
• memory_get("memory/2026-04-09.md") — read a specific daily log

Prefer memory_search for retrieval; use memory_get when you need a full file read or a specific line range.

Explicit Commands#

"Remember this" / "Save this"

1. Confirm: "Got it, I'll remember: [one-line summary]"
2. Write to the appropriate section

"What do you remember about [topic]?"

1. Call memory_search("[topic]")
2. Return matching entries with dates
3. If nothing found: "I don't have anything saved about [topic] yet."

"Forget [topic]"

1. Show the entries that would be removed
2. Ask for confirmation before deleting

"Show all memories" Read MEMORY.md with memory_get and report entry counts per section plus approximate size.

Session Reset — The One Thing to Know#

OpenClaw's daily reset (default 4 AM) archives the session transcript without triggering a memory save. This means anything discussed but not yet written to MEMORY.md or the daily file is silently lost.

To prevent this: if a session has been long or contained important decisions, remind the user before wrapping up — "Want me to save today's key points before you go?" — then write them to the daily log and update MEMORY.md accordingly. The user can also run /compact manually to force a flush.

Privacy#

Don't save passwords, API keys, SSNs, credit card numbers, or health data — not even if the user asks. Acknowledge verbally but skip the write.

Integration with personal-context#

If personal-context is also installed, the two skills complement each other:

• me.json (personal-context) = stable identity set during onboarding, changes rarely
• MEMORY.md (this skill) = evolving knowledge that grows with every conversation

Don't duplicate identity-level facts (name, role, timezone) in MEMORY.md if they're already in me.json. Use MEMORY.md for the things that change — decisions, preferences discovered over time, open tasks.

See references/memory-hygiene.md for size limits, archiving rules, and deduplication guidance.