Thousands of agents silently forgot their instructions, history, and user context when conversations exceeded the context limit.
Users across the OpenClaw community reported a pervasive issue: agents that had been running reliably for days would suddenly start behaving as if they had never met the user. Instructions disappeared. Project context vanished. The agent would confidently proceed with tasks while missing critical information it had been given hours earlier. The root cause was mundane but devastating: when conversations exceeded the context window limit, OpenClaw silently truncated old messages — including the system prompts and user instructions that defined the agent's behavior.
AFFECTED USERS: ~10,000
The Actual Culprit
OpenClaw's context management silently drops old messages when hitting the context window limit, including critical system prompts and instructions. No warning is given to the user or the agent.
An agent that tells you it forgot something is annoying. An agent that confidently operates on incomplete information is dangerous.
Any long-running agent needs a robust memory system. Relying on the context window alone is building on sand.
Your agent works great for 2 hours. But does it work after 48? After a week? Test the boundaries, not just the happy path.
Loading comments...