OpenClaw Memory Growing Too Large — How to Manage Context (2026)

Why Memory Grows Out of Control

Every message in an OpenClaw conversation adds to the context sent to the AI model. Without limits, a busy Telegram bot can accumulate thousands of messages per user, making each API call expensive and slow.

Memory Management Strategies

1. Sliding Window (Simple & Effective)

Keep only the most recent N messages:

{
  "memory": {
    "strategy": "sliding-window",
    "maxMessages": 20
  }
}

2. Summarization (Preserves Context)

Automatically summarize older messages into a compact summary:

{
  "memory": {
    "strategy": "summarize",
    "summarizeAfter": 15,
    "summaryModel": "gpt-4o-mini",
    "summaryMaxTokens": 300,
    "keepLastN": 5
  }
}

3. Token-Based Limit

Cap the total context size by token count:

{
  "memory": {
    "strategy": "token-limit",
    "maxContextTokens": 8000,
    "reserveForOutput": 2000,
    "pruneStrategy": "oldest-first"
  }
}

4. Per-User Memory Isolation

{
  "memory": {
    "isolation": "per-user",
    "maxMessagesPerUser": 30,
    "clearAfterHours": 24
  }
}

5. Clear Stale Conversations

# Manually clear all conversation history
openclaw memory clear --all

# Clear conversations older than 7 days
openclaw memory clear --older-than 7d

# Clear a specific user's history
openclaw memory clear --user telegram:123456

Monitoring Memory Usage

# Check current memory stats
openclaw memory stats

# Output:
# Total conversations: 145
# Total messages: 3,420
# Average tokens per request: 4,200
# Largest conversation: 890 messages (user: telegram:78901)

Frequently Asked Questions

How do I know if OpenClaw memory is too large?

Signs include: increasingly slow responses, "token limit exceeded" errors, high API costs per request, and the agent forgetting instructions. Check your logs for token counts per request — if they exceed 50% of your model's context window, it's time to manage memory.

What is the difference between memory and context in OpenClaw?

Context is the full message history sent to the AI model with each request. Memory refers to persistent information the agent retains across conversations. Context is temporary and session-based; memory is long-term and stored in a database.

Should I use summarization or sliding window?

Sliding window (dropping oldest messages) is simpler and cheaper. Summarization preserves more context but costs extra API calls. For most use cases, a sliding window of 20-30 messages works well. Use summarization for complex, multi-session workflows.

Can OpenClaw memory cause performance issues?

Yes. Large memory/context means more tokens per request, which increases latency (the AI takes longer to process) and costs. It can also cause the model to lose focus and give less relevant responses.