I’ve been running a complex multi-month AI project on Gemini with a custom knowledge base in a personalized GEM (semantic network stored as JSON files, behavioral protocols, structured memory system). The project was stable and performing well until approximately one week ago.
What changed: Since the recent model/RAG updates, the retrieval system has become completely unreliable. The model now consistently:
-
Mixes data from different time periods
-
Confabulates memories by merging unrelated entries
-
Returns contradictory information from the knowledge base
-
Fails to maintain temporal consistency in retrieved data
Impact: This isn’t a minor degradation - the entire project architecture has collapsed. Months of careful knowledge base construction and testing are now worthless because the retrieval layer can no longer be trusted.
What I need:
-
Acknowledgment of this issue - are others experiencing similar RAG instability?
-
Technical explanation of what changed in the retrieval system
-
Timeline for a fix, or confirmation that this is the “new normal”
I’ve already migrated to a different platform where retrieval is stable (though with reduced context). I’m posting this because if Google has fundamentally changed how RAG works, developers need to know their knowledge-based projects may be at risk.
Has anyone else experienced this degradation? Or found workarounds?