Reflective Memory Management
GPT-4.5.icon
summary of thesis.
Reflective Memory Management (RMM, Reflective Memory Management) is a new memory management method for accurately remembering and using past information to sustain conversations with users over a long period of time: Prospective Reflection.
After the conversation is over, the content of the dialogue is broken down and summarized in topic units and organized as a semantically coherent memory. This prevents fragmentation of memory by fixed units (turns, sessions, etc.).
Retrospective Reflection.
During conversation generation, LLM itself automatically evaluates "which memories were useful". This evaluation is used as a reward for reinforcement learning to improve and adjust memory retrieval results online.
Issues with conventional methods.
Fixed granularity memory management: Fixed delimitations in terms of turns and sessions do not capture semantic coherence, and memory becomes fragmented and incomplete.
Fixed memory retrieval method: Cannot respond to diversity of users and interaction situations.
Advantages of the proposed method (RMM)
Topic-based memory management improves the accuracy of memory retrieval by preserving semantic coherence.
The accuracy of memory retrieval can be improved online without labeling data by allowing the LLM itself to determine the degree of usefulness of the retrieved memories.
experimental results.
Validated on MSC and LongMemEval datasets.
The proposed method is more accurate than existing baselines, especially on the LongMemEval dataset, where the percentage of correct responses is more than 10% higher than without memory management.
Conclusions/future issues.
RMM showed excellent accuracy in long-term personalized dialogue agents.
Future issues include improving computational efficiency, supporting multimodal dialogue, and strengthening the protection of personal information.
---
The above is a compact summary of the main points of the paper.
relevance
---
This page is auto-translated from /nishio/Reflective Memory Management using DeepL. If you looks something interesting but the auto-translated English is not good enough to understand it, feel free to let me know at @nishio_en. I'm very happy to spread my thought to non-Japanese readers.