On the same day, the minimum number of rounds 4 previously discussed was 4, which can be extracted for memory. However, later in the conversation, without making any modifications, and with no indication in the reasoning log that the extraction conditions were not met, memory extraction was still not possible. We continued the conversation for over a dozen entries, and I kept changing the prompt words, the minimum number of rounds, switching models, and trying different memory extraction prompts, but none were successful. The default prompt words didn’t work either. I tried using the models gemini3 and DeepSeekV3.2. Is it an issue with my settings or is there a bug in the automatic extraction? I accidentally leaked the key, so I generated a new one.
I need to extract memories in real-time for voice call gameplay to update topics; otherwise, the conversation won’t connect. Previously, I used permanent memories and then added new topic summaries.
I’m not sure if this is the reason (it was the cause when I encountered a similar issue): The minimum number of memory retrieval dialogues should be set to twice the number of rounds that trigger memory retrieval.
It shouldn’t be this 12 14 24 44, tried all sorts of them, sometimes after replying there’s a noticeable loading, feels like it’s extracting, but nothing gets extracted.
When it comes to the trigger round, check the reasoning log for any formatting errors.
Also, you need to pay attention to the default memory extraction prompt: Note: If the information is duplicated or conflicts with known information, ignore this information. <Known Information>{{memories}}</Known Information> This section will make the summary model compare the extracted memories, and duplicate information will not be extracted.
This topic was automatically closed 3 hours after the last reply. New replies are no longer allowed.