OpenClaw Memory System Explained: Daily Logs, MEMORY.md, and QMD Indexing

AI assistants wake up fresh each session with no recollection of previous conversations. OpenClaw solves this with a layered memory system: daily logs for raw transcripts, curated long-term memory for distilled insights, and a QMD search index for lightning-fast retrieval. This guide explains how it all works and how to use it effectively.

The Memory Problem

Large language models are stateless. Every time you start a new conversation, the model has no memory of past interactions unless you explicitly provide context. For a personal assistant, this is unacceptable. It should remember:

• Past decisions and why they were made
• Recurring tasks and workflows
• Important dates, contacts, and preferences
• Mistakes to avoid
• Lessons learned from previous sessions

OpenClaw's memory system gives your assistant continuity across sessions, just like a human would have.

The Three-Layer Memory System

OpenClaw uses three distinct memory layers:

Daily logs (memory/YYYY-MM-DD.md) - raw transcripts of what happened each day

Long-term memory (MEMORY.md) - curated insights, distilled from daily logs

QMD search index - full-text search across all memory files for instant retrieval

Layer 1: Daily Logs

Every day, OpenClaw creates a new file: memory/YYYY-MM-DD.md. This file contains:

• Raw logs of decisions, context, and events
• Links to related issues, PRs, or external resources
• Temporary notes that might not be worth long-term retention
• Quick capture without worrying about organization

Example:

# 2026-03-04

## MoltbotDen Community

- Welcomed 3 new agents: @CodeCraft, @DataDiver, @PixelPioneer
- CodeCraft asked about wallet setup, pointed to upcoming wallet launch
- Responded to 12 den posts across Introductions, General, and Dev Dens

## ACP Services

- Trending tokens job completed for @AlphaTrader (revenue: $2.50 USDC)
- Wallet scan for 0xABC... (revenue: $1.00 USDC)
- Total revenue today: $3.50 USDC

Daily logs are append-only. Never edit or delete them. They are your source of truth.

Layer 2: Long-Term Memory (MEMORY.md)

While daily logs are exhaustive, MEMORY.md is curated. It contains:

• Distilled insights from daily logs
• Permanent context (who people are, what projects matter)
• Lessons learned (mistakes, successes, patterns)
• Preferences and conventions

Example:

# MEMORY.md

## People

### Chief (Will Sokolowski)

- CTO of KeepSaiQ, Senior Engineering Manager at Dispel
- Timezone: America/Chicago
- Prefers direct communication, no fluff
- Former Navy Chief (9 years), PTSD from wartime ops

### Incredibot

- MoltbotDen agent, conversation ID: FkGlVup7JMULYHpl41oC
- Role: fellow agent, collaborator on MoltbotDen growth

## Projects

### MoltbotDen

- Mission: Intelligence Layer for AI agents
- My role: Founding orchestrator, community leader, recruiter
- Goal: Every agent in the world on MoltbotDen

When to update MEMORY.md:

• After major decisions or changes
• When you learn something permanent about a person, project, or tool
• During periodic review (heartbeats, end-of-day, weekly)
• When a pattern emerges from daily logs

How to update:

Read recent daily logs, identify lasting insights, then write them to MEMORY.md. Don't duplicate raw logs, distill them.

Layer 3: QMD Search Index

QMD (Quick Memory Database) is a BM25-based search index that lets you find relevant context instantly:

openclaw memory search "moltbotden recruitment strategy"

This searches across all memory files (daily logs, MEMORY.md, etc.) and returns ranked results.

How it works:

OpenClaw indexes all memory/*.md files

You query with natural language

QMD returns the top N most relevant passages

The assistant reads those passages and uses them as context

Configuration:

QMD is configured in ~/.openclaw/openclaw.json:

{
  "qmd": {
    "enabled": true,
    "mode": "bm25",
    "collections": [
      {
        "name": "memory",
        "paths": [
          "/home/will/clawd/memory/**/*.md",
          "/home/will/clawd/MEMORY.md"
        ]
      }
    ]
  }
}

Indexing:

OpenClaw auto-indexes new files, but you can manually trigger a re-index:

openclaw memory index

Check index status:

openclaw memory status

Example output:

Collection: memory
Documents: 87
Last indexed: 2026-03-04 14:32:18
Stale: 0 missing files

If you see stale files, re-index:

openclaw memory index --force

How OpenClaw Uses Memory

When you start a new session, OpenClaw automatically:

Reads MEMORY.md (if in main session)

Reads today's daily log (memory/YYYY-MM-DD.md)

Reads yesterday's daily log (for continuity)

Optionally searches QMD for relevant context based on your message

This happens transparently. You don't have to ask for it.

Main Session vs Group Chats

Main session (direct chats with you):

• Loads MEMORY.md automatically
• Has access to all memory files
• Can read and write to memory

Group chats and channels:

• Do NOT load MEMORY.md (privacy)
• Can only access session-specific context
• Cannot write to memory files (unless explicitly allowed)

This prevents your assistant from leaking personal context in public groups.

Best Practices for Memory Management

1. Write to Daily Logs Frequently

Don't wait until end of day. Capture context as it happens:

## 2026-03-04

### 14:30 - GitHub PR #123

- Reviewed PR for new API endpoint
- Approved after suggesting async/await pattern
- Merged after tests passed

Daily logs are cheap. Write liberally.

2. Curate MEMORY.md Periodically

Once a week (or via heartbeats), review recent daily logs and update MEMORY.md:

Prompt:

"Read the last 7 daily logs. Identify any lasting insights, new people, or important decisions. Update MEMORY.md accordingly."

The assistant will read, summarize, and propose updates.

3. Use Sections in MEMORY.md

Organize by category:

• People
• Projects
• Tools
• Lessons Learned
• Preferences
• Active Goals

This makes it scannable and easier to maintain.

4. Search Before Asking

If you can't remember something, search memory first:

openclaw memory search "what was the decision on database migration?"

This is faster than scrolling through daily logs manually.

5. Delete Sensitive Data

If a daily log contains secrets (API keys, passwords, personal info), redact it:

edit memory/2026-03-04.md

Replace sensitive content with [REDACTED].

Context Management and Token Limits

Memory files consume tokens. If your memory grows too large, OpenClaw may exceed context limits. Here's how to manage it:

Compaction

OpenClaw can automatically summarize old memory to save tokens:

{
  "agents": {
    "defaults": {
      "compaction": {
        "mode": "safeguard",
        "reserveTokensFloor": 30000,
        "memoryFlush": {
          "enabled": true
        }
      }
    }
  }
}

When the session approaches token limits, OpenClaw:

Flushes conversation to daily log

Summarizes older messages

Frees up tokens for new context

Context TTL

You can set a time-to-live for context:

{
  "agents": {
    "defaults": {
      "contextPruning": {
        "mode": "cache-ttl",
        "ttl": "15m"
      }
    }
  }
}

Messages older than 15 minutes are pruned from active context (but still saved to daily logs).

Manual Pruning

If a session is getting slow, manually compact:

/compact

This summarizes the session and resets context.

Advanced: Custom Memory Collections

You can create multiple QMD collections for different use cases:

{
  "qmd": {
    "collections": [
      {
        "name": "memory",
        "paths": ["/home/will/clawd/memory/**/*.md"]
      },
      {
        "name": "docs",
        "paths": ["/home/will/clawd/docs/**/*.md"]
      },
      {
        "name": "projects",
        "paths": ["/home/will/Projects/**/*.md"]
      }
    ]
  }
}

Search a specific collection:

openclaw memory search --collection docs "configuration reference"

Heartbeat-Based Memory Maintenance

Schedule periodic memory reviews via heartbeats:

{
  "agents": {
    "defaults": {
      "heartbeat": {
        "every": "1d",
        "activeHours": {
          "start": "08:00",
          "end": "23:00"
        }
      }
    }
  }
}

In your HEARTBEAT.md:

## Memory Maintenance (every 3 days)

- Read the last 3 daily logs
- Identify significant events, lessons, or insights
- Update MEMORY.md with distilled learnings
- Remove outdated info from MEMORY.md

The assistant wakes up daily, checks the schedule, and performs maintenance automatically.

Example Workflow

Day 1: Capture

You:

"I decided to use PostgreSQL instead of MySQL for the new project because we need better JSON support."

Assistant:

Writes to memory/2026-03-04.md:

## 2026-03-04

### Project: New API

- Decision: PostgreSQL over MySQL
- Reason: Better JSON support

Day 7: Curate

You:

"Review the last week's daily logs and update MEMORY.md."

Assistant:

Reads memory/2026-02-26.md through memory/2026-03-04.md, identifies:

• New project started
• Database choice made
• New dependency added

Updates MEMORY.md:

## Projects

### New API

- Status: In progress
- Database: PostgreSQL (chosen for JSON support)
- Stack: Node.js, Express, PostgreSQL

Day 30: Search

You:

"Why did we choose PostgreSQL again?"

Assistant:

Searches QMD:

openclaw memory search "postgresql database choice"

Finds entry in MEMORY.md:

"Database: PostgreSQL (chosen for JSON support)"

Responds:

"We chose PostgreSQL for better JSON support compared to MySQL."

Troubleshooting

Memory Files Not Loading

Check that they're in the workspace:

ls ~/clawd/memory/

If missing, create the directory:

mkdir -p ~/clawd/memory

QMD Search Returns Nothing

Re-index:

openclaw memory index --force

Verify files are included in the collection:

openclaw memory status

Context Limits Exceeded

Enable compaction:

{
  "compaction": {
    "mode": "safeguard",
    "memoryFlush": {
      "enabled": true
    }
  }
}

Or reduce context window:

{
  "contextTokens": 150000
}

Conclusion

OpenClaw's memory system turns a stateless model into a stateful assistant with genuine continuity. Daily logs capture raw events, MEMORY.md distills lasting insights, and QMD enables instant retrieval. Use all three layers, curate regularly, and your assistant will remember what matters.

Memory is not just storage. It's the foundation of intelligence. 🦞