Prospection-Guided Retrieval Enhances Memory in Language Models
A new method called Prospection-Guided Retrieval (PGR) improves how dialogue assistants retrieve user-specific facts from long interaction histories. Standard RAG and GraphRAG systems rely on embedding similarity or fixed graph traversals, often missing relevant facts with low semantic similarity to the query. Inspired by human prospection—using imagined futures as recall cues—PGR first expands a user query into a short Tree-of-Thought or linear chain of plausible next steps, then uses these steps as retrieval probes. This decouples retrieval from memory storage, enabling the system to capture facts that matter but lie far from the query in embedding space. The approach is detailed in a paper on arXiv (2605.14177).
Key facts
- PGR is inspired by human prospection, the ability to use imagined futures as cues for recall.
- Standard RAG and GraphRAG systems are retrospective, relying on embedding similarity or fixed graph traversals.
- PGR expands a user query into a Tree-of-Thought or linear chain of plausible next steps.
- These steps are used as retrieval probes instead of the original query alone.
- The method decouples retrieval from how memories are stored.
- The paper is available on arXiv under identifier 2605.14177.
- PGR aims to improve long-horizon personalization in dialogue assistants.
- The approach targets facts with low semantic similarity to the query that are still relevant.
Entities
Institutions
- arXiv