MemChain: A Trainable Memory Policy for LLM Agents
MemChain serves as a trainable memory policy designed for memory-augmented LLM agents post-retrieval. It converts retrieved candidates into an active memory focused on answers, encapsulated as a compact and grounded evidence context. When presented with a user query and the retrieved candidates, MemChain formulates a question-conditioned evidence plan, creating an ordered trace of grounded evidence that categorizes memories based on semantic roles and dependencies. It also performs explicit memory actions to yield a succinct evidence context for generating answers. This method overcomes the challenges posed by the retrieval-as-evidence framework, which presumes the appropriateness of retrieved memories for reasoning, often leading to redundancy, conflicts, and weak relevance, thereby increasing context overhead in long-term memory tasks.
Key facts
- MemChain is a trainable post-retrieval memory policy.
- It transforms retrieved candidates into answer-facing active memory.
- It generates a question-conditioned evidence plan.
- It constructs an ordered grounded evidence trace.
- It executes explicit memory actions to produce a concise evidence context.
- It addresses the retrieval-as-evidence paradigm's limitations.
- The retrieval-as-evidence paradigm assumes retrieved memories are suitable for reasoning.
- The paradigm incurs substantial context overhead in long-term memory tasks.
Entities
—