Anyone else seeing GPT-5.2 not using saved memories unless explicitly asked?

Exploring Memory Functionality in GPT-5.2: Are Saved Memories Being Utilized Effectively?

Recent observations indicate that GPT-5.2 exhibits some inconsistencies with its handling of stored memories during interactions. Users have reported that, despite having persistent memories saved, the AI often does not incorporate these memories into ongoing conversations unless explicitly prompted.

Understanding Persistent Memory in GPT-5.2

GPT-5.2 offers the capability to save reference memories, intended to enhance contextual awareness across sessions. However, current behaviors suggest that these stored memories are not automatically integrated into typical interactions. Instead, the model responds primarily based on the immediate chat history, requiring users to specifically ask about their saved memories to retrieve relevant information.

Examples of Observed Behavior

• When requesting a personalized summary, such as “2025 wrapped,” responses are limited to the current conversation, with no reference to previously saved details.

• Discussions about travel plans or future trips tend to ignore countries already visited, despite having these details stored in memory.

• Asking direct questions like “Which countries have I visited?” correctly fetches the stored information, indicating that the memory itself is intact but not automatically used.

• Reference to saved chat history appears to be non-functional; the model does not spontaneously incorporate prior information without prompt.

Device and Platform Considerations

Interestingly, users have noted that the desktop application on macOS maintains more consistent memory behavior compared to the web client. Re-logging out, clearing browser cache, and restarting sessions can sometimes mitigate issues on the web platform, suggesting the root cause may lie within the session or cache management rather than the memory system itself.

Implications for Users

This discrepancy raises questions about the practical utility of persistent memory in GPT-5.2. If stored memories require manual activation, their effectiveness as a contextual assistant diminishes, transforming memory from an integrated feature into a manual database lookup.

Next Steps and Community Input

Current reports are primarily based on user experiences and observations, which are valuable for understanding the system’s behavior. It is important to distinguish between actual technical limitations and potential user-side issues, especially given that responses about system internals are often speculative.

For users experimenting with GPT-5.2, awareness of this behavior can inform how they structure interactions to maximize the utility of saved memories.

Conclusion

While GPT-5.2’s persistent memory feature holds promise for creating more personalized interactions, its inconsistent automatic utilization limits its potential. Continued community discussion and official updates may clarify whether this is a bug