openclaw - 💡(How to fix) Fix [Bug]: Anthropic Prompt Cache Invalidation on Startup Context (7x cost increase)

Bug type

Behavior bug (incorrect output/state without crash)

Beta release blocker

No

Summary

OpenClaw auto-loads dynamic memory files at startup, breaking Anthropic's prompt cache and causing 7x cost increase

Steps to reproduce

1. Set up OpenClaw with Anthropic Claude API
2. Use persistent memory files (MEMORY.md, current_state.json, current_events.md)
3. Start a session and send a message
4. Check Anthropic console → token usage is 8x higher than expected
5. Modify any memory file
6. Send another message → cache hit stays 0%

Expected behavior

Prompt cache should hit 80-90%, keeping per-message cost low (~$0.002-0.01 per short reply)

Actual behavior

Prompt cache hit 0%, every message costs $0.08+ (8x normal). Burn rate: $45/day instead of $6-10/day

OpenClaw version

OpenClaw 2026.5.7 (eeef486)

Operating system

Macos

Install method

Npm openclaw onboard

Model

Claude Haiku + Opus (for synthesis)

Provider / routing chain

Anthropic Claude API

Additional provider/model setup details

No response

Logs, screenshots, and evidence

Impact and severity

CRITICAL - $1,080-1,200/month waste. Makes OpenClaw unusable with Anthropic for persistent memory users.

Additional information

Root cause: Startup context auto-loads MEMORY.md, current_state.json, current_events.md. These files change multiple times per day. Each change = OpenClaw re-injects = Anthropic cache invalidation.

Evidence:

• May 15 04:51: $52.97
• May 15 04:54: $52.72 (after 3 exchanges)
• $0.25 burned in 3 minutes = $0.08 per reply (should be ~$0.003)

Proposed fix: Config option to exclude dynamic files from auto-load: startup: excludePatterns: - memory/MEMORY.md - memory/current_.json - memory/current_.md