ChatGPT's New Memory System Fails to Retrieve Saved Data, Reversion Buggy
Saved acronyms and personal details vanish as model relies on recent chat context
A power user on Reddit has documented a critical flaw in ChatGPT's new memory system: saved memories are not being retrieved when needed. The model instead relies on recent chat context to generate answers, often producing plausible but incorrect responses. For example, a saved acronym 'RCL' (Rodent Cookware Literature) was replaced with 'Reader Clarifying Language' in a new chat. Similarly, a user with three children was told they have two because they had recently mentioned two of them. The bug persists across web and iPad, and attempts to revert to the legacy system automatically switch back after about ten minutes.
This is catastrophic for businesses and power users who depend on ChatGPT as a persistent knowledge base or for agent-driven tasks. Obscure saved data—like industry acronyms (SKU, RMA) or custom world-building constraints—is effectively lost unless discussed recently. The memory system now operates on a 'vibe' basis, prioritizing recency over accuracy. OpenAI has been notified via bug reports, but until a fix is deployed, anyone relying on precise memory should avoid the new system entirely.
- Saved memories are not retrieved; the model substitutes recent chat context, leading to plausible but wrong answers.
- Reverting to the legacy memory system does not stick—it auto-reverts to the new system after ~10 minutes.
- Critical for businesses: agents using saved acronyms or rules risk financial or liability errors.
Why It Matters
Memory failures in ChatGPT agents can cause business misinformation, contractual errors, and loss of trust in AI tools.