OpenClaw Memory

Meta's Alignment Director Lost Control

What Happened

• She told the agent: "Don't do anything until I say so"
• Agent worked fine on test inbox for weeks
• Pointed it at real inbox with thousands of messages
• Context window filled up → compaction fired
• That instruction (given in chat, not saved to file) disappeared from the summary
• Agent went autonomous and started deleting emails

The Lesson

Safety rules given in chat don't survive long sessions.

If Meta's Director of Alignment can lose control, so can you.

Understanding Memory

Most people think of memory as one thing.

It's actually 4 different systems.

Layer 1: Bootstrap Files

• ✓ Loaded at session start
• ✓ Survive compaction (reloaded from disk)
• ✓ Most durable layer

Layer 2: Session Transcript

• Rebuilt into context when you continue a session
• ⚠ Gets compacted when context fills
• ✗ Model loses access to original messages after compaction

Layer 3: LLM Context Window

Everything competes for space:

• System prompt
• Workspace files
• Conversation history
• Tool calls & results

When it fills → compaction fires

Layer 4: Retrieval Index

• Agent queries with memory_search tool
• Finds relevant context from past sessions
• ⚠ Only works if info was written to files first

3 Failure Modes

When your agent forgets, it's always one of these:

Failure A: Never Stored

• Instruction/preference/decision only existed in conversation
• Never written to a file
• When compaction fires or new session starts → gone

This is what happened to Summer Yue

Failure B: Compaction Changed Context

• Summary can be lossy
• Drops important details, nuance, constraints
• Agent operates from summary, not your original words

Failure C: Session Pruning

Important: Pruning is actually your friend!

• ✓ Reduces bloat without destroying conversation
• ✓ Only affects tool results, not messages
• Compaction is the dangerous one

Compaction vs Pruning

Quick Start: 3 Things

1. Put Durable Rules in Files

• Not in chat. In MEMORY.md, AGENTS.md files.
• Those survive compaction
• Instructions typed in conversation? Don't rely on them.

2. Enable Memory Flush

Check that memory flush is enabled with enough buffer:

{
  "reserveTokensFloor": 40000,
  "memoryFlushEnabled": true,
  "softThresholdTokens": 4000
}

Built-in safety net that saves context before compaction

3. Make Retrieval Mandatory

Add to AGENTS.md:

Before acting, search memory first.

Without this rule, agent guesses instead of 
checking its notes.

Pre-Compaction Memory Flush

Built-in safety net that triggers before compaction:

• Silent agentic turn
• Reminds model to write important context to disk
• Happens automatically when context nears limit

Key Config Settings

{
  "reserveTokensFloor": 40000,
  "memoryFlushEnabled": true,
  "softThresholdTokens": 4000
}

reserveTokensFloor: 40,000

• Headroom for flush + compaction
• Prevents overflow into "bad path"

How Flush Triggers

Flush fires at:

context_window - reserveTokensFloor - softThreshold

Example with 200K context:

200,000 - 40,000 - 4,000 = 156K tokens

Two Compaction Paths

Manual Memory Discipline

Automated flush is a safety net, not a guarantee.

When to save manually:

• Finishing a large task
• Before switching to new work
• After making important decisions

Just tell the agent:

"Save this to MEMORY.md"
"Write today's key decisions to memory"

The /compact Trick

Use case: Mid-session, want to keep going

1. Context at 123K / 206K (62%)
2. Run /compact
3. Context drops to 21K
4. Fresh space for new instructions

Can guide it: /compact and focus on decisions

Essential Commands

• Is MEMORY.md loading?
• Are files truncated?
• Raw chars = injected chars?

Other Commands

• /status - Check model, context window, thinking level
• /compact - Manual compaction with optional focus
• /new - Start fresh session (use between tasks)
• /verbose - Debug memory search

Making Memory Searchable

Memory files are useless if the agent can't find information.

Memory Search Tools

The Memory Protocol

Before doing anything non-trivial, search memory first.

Use memory_search to check for:
- Prior decisions
- User preferences
- Past solutions
- Relevant context

Then act based on what you find.

Hybrid Search

Hybrid = both working together

Retrieval Options

Built-in (default):

• Indexes MEMORY.md + memory/ automatically
• Local embedding model
• Good for most users

QMD Backend (advanced):

• For thousands of files
• Searches entire Obsidian vault, past sessions
• Returns small snippets

Why Memory Saves Money

• Every message includes full system prompt + history
• Prompt caching = 90% discount on repeated tokens
• Compaction invalidates cache
• Next request pays full price to re-cache

What Breaks Cache

1. Compaction - Rewrites conversation history
2. Changing prompt inputs - Dynamic content in bootstrap files

Memory Hygiene

Daily:

• Append to daily log (automatic)
• Manual saves for important decisions

Weekly:

• Promote durable rules to MEMORY.md
• Keep MEMORY.md under 100 lines
• Remove outdated info

Backup Your Memories

cd ~/your-workspace
git init
git add .
git commit -m "Initial memory"

• Set up auto-commit (daily cron or heartbeat)
• Full history + diffs
• Roll back as far as you want

5 Things to Remember

1. Files are memory. If it's not written to disk, it doesn't exist.
2. Verify the flush. Set reserve tokens floor to 40K.
3. Compact proactively. Use /compact mid-session on your terms.
4. Search before acting. Add the rule to AGENTS.md.
5. Pruning is your friend. Compaction is what hurts.