基于隐式记忆构建的检索生成方法、装置及电子设备
By constructing the implicit memory state of the agent through multimodal rendering and visual encoding, and combining progressive memory solidification and variational implicit memory encoding, the problem of balancing memory compression efficiency and semantic fidelity in long-term interactions of the agent is solved, achieving efficient query statement retrieval and generation and avoiding performance degradation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BAIDU INT TECH (SHENZHEN) CO LTD
- Filing Date
- 2026-03-17
- Publication Date
- 2026-07-17
AI Technical Summary
In existing technologies, intelligent agents struggle to balance memory compression efficiency and semantic fidelity during long-term interactions, leading to semantic shifts and information loss, performance degradation, and impacting the efficiency and accuracy of query generation.
By acquiring historical interaction information of the agent, performing multimodal rendering and visual encoding, constructing a visual semantic vector of implicit memory state, generating retrieval results using a target autoregressive decoder, and combining progressive memory solidification mechanism and variational implicit memory encoding, efficient compression of memory and semantic fidelity are achieved.
Achieving a balance between memory compression efficiency and semantic fidelity reduces semantic shift and information loss, prevents performance degradation of agents in long-term interactions, and improves the efficiency and accuracy of query generation.
Smart Images
Figure CN122412440A_ABST