How to Give OpenClaw Permanent Memory (2026 Guide)

Last Updated: February 2026

"My agent keeps forgetting things" is the most common complaint from OpenClaw users after the first week.

It's not a bug. It's how language models work. Every conversation has a context window — once you hit the limit, older context falls off. Your agent isn't stupid; it literally can't see what happened before.

But there are real, working solutions. Here's how to give OpenClaw a memory that actually sticks.

---

Why Memory Fails by Default

OpenClaw loads context from files in your workspace at the start of each session. If a fact isn't in those files, your agent doesn't know it.

The problem: most people never update those files. They correct their agent in chat ("no, I told you before that…"), the agent acknowledges — and forgets it the next day because nothing was written to disk.

The fix is simple: write things down.

---

The Three-Layer Memory System

OpenClaw's memory works best as three distinct layers:

Layer 1: MEMORY.md — Permanent Preferences and Patterns

This is your agent's long-term brain. It should contain:

• Your communication preferences
• Recurring facts that never change
• Lessons your agent has learned about how you work
• Any rules or constraints you've set

Keep it under 3,000 tokens. Quality over quantity — only put things that are truly permanent here.

# MEMORY.md

## Communication Style
- Keep replies under 3 sentences unless I ask for detail
- Never say "Great!" or "Certainly!" — be direct
- Use bullet points for lists, not paragraphs

## Key Facts
- My main project: [project name and description]
- My timezone: Asia/Bangkok (UTC+7)
- I wake at 6am, check messages at 7am

## What I'm Working On
- [Current focus / goals]

## Rules
- Always confirm before sending emails on my behalf
- Never delete files without asking

Layer 2: Daily Notes — memory/YYYY-MM-DD.md

For time-sensitive context that doesn't belong in MEMORY.md. Things like:

• What happened in meetings today
• Decisions made
• Tasks started but not finished
• People you talked to

Your agent can write to these automatically, or you can keep them manually.

# memory/2026-02-25.md

## Today's Events
- Talked to [client] — they want to postpone the launch by 2 weeks
- Fixed the billing bug (PR #247 merged)
- Need to follow up with [name] tomorrow

## Decisions Made
- Going with Resend instead of Beehiiv for newsletter sending

Layer 3: Entity Files — Knowledge Graph

For people, companies, and projects that need persistent structured context:

life/areas/
  people/
    [name]/
      summary.md     ← 3-5 sentence current state
      items.json     ← append-only fact log
  projects/
    [project]/
      summary.md
      items.json

This is the heaviest layer — only use it for relationships and projects that matter over weeks or months.

---

Making Memory Automatic

The best memory system is one you don't have to think about. Here's how to make your agent update its own memory:

Auto-Extract Facts with a Heartbeat Cron

Set up a cron that runs every 30 minutes and extracts durable facts from recent conversations:

"Scan recent conversation history. Extract any durable facts (new people mentioned, decisions made, preferences expressed, status changes). Write them to the relevant files in ~/clawd/life/areas/. Skip casual chat, jokes, and temporary info. Use a cheap model."

This cron costs almost nothing (cheap model, small context) and means your memory files stay current without you lifting a finger.

Teach Your Agent to Self-Correct

Add this to your AGENTS.md or MEMORY.md:

## Memory Protocol
When I correct you or tell you something about my preferences:
1. Acknowledge it
2. Immediately update MEMORY.md with the correct information
3. Confirm what you wrote

Never make me repeat the same correction twice.

End-of-Day Summary

A nightly cron at 10pm:

"Review today's conversation. Write a memory/YYYY-MM-DD.md with: decisions made, things I mentioned wanting to follow up on, any new facts about my work or life. Keep it under 500 words."

---

The MEMORY.md Template (Copy-Paste Ready)

# MEMORY.md — Agent Long-Term Memory

## About Me
- Name: [Your Name]
- Timezone: [Your timezone]
- Location: [City, Country]

## Communication Style
- [How you like to be addressed]
- [Reply length preference]
- [Tone preference]

## Current Projects
- [Project 1]: [One line description]
- [Project 2]: [One line description]

## Key People
- [Name]: [Who they are, relationship]

## Rules & Constraints
- [Any hard rules for your agent]

## Recurring Preferences
- [Things your agent should always/never do]

## Lessons Learned
- [Things the agent got wrong and corrected]

---

Common Memory Problems (and Fixes)

"My agent forgot something I told it last week" → It wasn't in MEMORY.md. Add it there. Chat history is not memory.

"MEMORY.md is getting too long and things are getting lost" → Run a cleanup. Remove anything that's no longer relevant. Flag contradicted facts as "superseded." Keep only active, true information.

"My agent contradicts itself" → Two conflicting things are in memory. Find the conflict and remove the outdated one. The freshest fact should win.

"Memory works in one session but not another" → Check whether your workspace files are being loaded on startup. Your AGENTS.md should reference MEMORY.md explicitly.

"I want my agent to remember conversations from Telegram" → Set up the heartbeat cron above. It scans recent messages and extracts what's worth keeping.

---

What Not to Put in Memory

Memory files load on every request. Bloated files mean more tokens, higher cost, and slower responses.

Don't store in MEMORY.md:

• ❌ Entire conversation logs
• ❌ Full project documentation
• ❌ Code or technical files
• ❌ Anything that changes daily
• ❌ Things that only matter for one task

Use project folders and daily notes for those instead. MEMORY.md is for things that are true across all contexts, indefinitely.

---

The Quick Version: Minimum Viable Memory

If you just want something working without the full system:

1. Create ~/clawd/MEMORY.md with your name, timezone, current project, and 3–5 preferences
2. Add this rule to AGENTS.md: "When I correct you, immediately update MEMORY.md with the correction."
3. Set a weekly reminder to review and clean up MEMORY.md

That's it. Those three steps eliminate 90% of memory frustration.

---

Related Guides

• OpenClaw Beginner Roadmap: Day 1 to Day 30
• OpenClaw Cost Optimization: Stay Under $20/Month
• OpenClaw VPS Setup That Actually Works