OpenClaw's memory is unreliable, and you don't know when it will break

Article Summary

Drawing on observations of roughly 1,000 OpenClaw deployments, the author argues there are effectively zero legitimate production use cases because the agent’s persistent memory is fundamentally unreliable — it forgets context mid-task in ways users cannot detect or correct. He frames this not as a fixable bug but as a core architectural constraint, concluding that OpenClaw currently functions as little more than a chatbot with extra steps.

Discussion

• Heavy skepticism dominates: multiple commenters echo that agent workflows just move the management burden from the real job to babysitting the AI
• A dissenting voice lists active team uses — SDR research, proposal drafting, staging ops — including a $40K proposal generated from meeting notes
• Memory and context loss are the central technical pain point: agents randomly edit their own config, use wrong JSON keys, and require constant verification
• Builders share workarounds — a belief-based SQLite system that supersedes stale facts, and Agent Kanban for VS Code — but all acknowledge an ~85% success rate is insufficient for autonomy

Discuss on HN