Evolving Agent Memory
Model memory as a changing body of knowledge with explicit operations for creation, revision, merging, and forgetting.
Current research
Infrastructure for
evolving memory.
How can an Agent's memory change continuously without losing consistency, efficiency, or the ability to forget?
Thesis direction
Versioned Multi-representation Storage and Utility-aware Scheduling for Evolving Agent Memory
面向演化 Agent 记忆的版本化多表示存储与效用感知调度研究
Versioned Representations
Track derivation and validity across text, summary, embedding, and model-specific execution artifacts.
KV Cache as a View
Treat KV cache as a materialized view bound to model, tokenizer, prompt layout, and execution environment.
Utility-aware Scheduling
Balance latency, quality, reuse probability, recomputation cost, and resource pressure across tiers.
Consistency & Forgetting
Propagate invalidation safely when factual sources are edited, merged, expired, or deleted.
System model
Facts at the center.
Views around them.
The factual source remains authoritative. Derived representations are versioned, traceable, and disposable. This boundary makes reuse safe and forgetting meaningful.
Execution viewKV Cachemodel · tokenizer · layout
Semantic viewsEmbedding · Summaryderived · versioned · replaceable
Factual sourceCanonical Memoryauditable · editable · authoritative