Retrieval as Traversal: Value-Indexed Memory for Stateful Long-Context Agents
What it is
This paper proposes a new memory architecture for long-context agents that treats retrieval as a persistent local traversal through an address space rather than a global search. It separates memory content from address, using a cost-regularized objective to balance information value against retrieval cost, which encourages complementary memories to be co-retrievable while excluding redundant ones.
The problem it solves
Current long-context models either repeatedly search for relevant information in every forward pass (attention) or decide too early what to preserve (recurrent compression), failing to exploit the overlap between useful memories at neighboring retrieval steps.
What you could use it for
- Designing memory systems for stateful agents that need to retain and retrieve information over long contexts
- Optimizing retrieval efficiency in recurrent or attention-based models by leveraging local neighborhood overlaps
- Developing cost-aware indexing strategies for large-scale knowledge bases
Why it matters
If successful, this approach could significantly reduce the computational cost of retrieval in long-context models by making stateful retrieval easier to learn than stateless global search, potentially enabling more efficient and scalable agent memory.
Read the source ↗Back to the library →
Summary written by an AI model from the source; check the source before you rely on it.