Context Too Large: Managing Session Memory and Avoiding Token Burns

Understanding the Problem

When you see "Context overflow: prompt too large for the model", your conversation history has exceeded the model's token limit. But there's a bigger issue: token burns.

What Causes Token Burns

The cost spike isn't because you "asked something big" — it's because big data was accidentally pulled into the main DM context, and then every normal message kept dragging that context along.

Common Culprits

Commands that returned huge outputs get stored in the session transcript:

openclaw config.schema → massive JSON schema

openclaw status --all → full system dump

Large file contents or log dumps

Verbose command outputs

After that, even small questions cause the model to process gigantic cached context.

Symptoms

Sudden spike in API costs

"Context overflow" errors

Session randomly resets

Agent "forgets" recent conversations

Slow response times

Immediate Fixes

1. Reset the Session

Stop carrying the giant context forward:

# In chat, simply say: /new

# Or via CLI: openclaw session reset

2. Check Current Context Size

openclaw session info

Look at the token count - if it's near your model's limit (200k for Claude), that's your problem.

Prevention Strategies

Configure Aggressive Compaction

{
  "agents": {
    "defaults": {
      "contextTokens": 50000,
      "compaction": {
        "mode": "safeguard"
      }
    }
  }
}

Safeguard mode automatically:

Detects when context approaches limits

Compacts the session before overflow

Recovers by truncating old messages if overflow occurs

Avoid Big Outputs in Main Session

Never run "big output" tools in your main DM:

# BAD: Dumps huge output into main context openclaw config.schema

# GOOD: Run in isolated debug session openclaw --session debug config.schema

Set History Limits

{
  "session": {
    "historyLimit": 50
  }
}

This keeps only the last 50 messages, preventing unbounded growth.

Understanding Memory vs Sessions

Sessions: Conversation history (can overflow)

Memory: Durable facts written to disk (persists forever)

When compaction happens, the agent should write important facts to memory before context is cleared. Enable this with:

{
  "agents": {
    "defaults": {
      "memory": {
        "enabled": true,
        "autoSave": true
      }
    }
  }
}

Monitoring Token Usage

Watch your token consumption:

# Real-time session monitoring
openclaw logs --follow | grep "tokens"