Purpose-Driven Associative Retrieval: Building a User-Modeling Context from a Memory Graph
Version note. v0.5.1 only adds the archived DOI statement to the first page; all results are unchanged from v0.5. Abstract. Retrieval research mostly targets question answering: given a query, return passages that contain the answer. Dialogue memory poses a different problem. The system must reconstruct the user at this moment — their state, preferences, relationships, constraints, and history — rather than locate a passage. This task has no stable golden set: the same memory may be relevant under one intent and irrelevant under another, so P@K and R@all cannot define “correct” or compare systems that build context differently. We present PAR (Purpose-driven Associative Retrieval), a set-valued retrieval mechanism over a typed memory graph. PAR expands candidates along typed relations with a direction-aware weight matrix, scores them by purpose relevance, and stops by a peak-tolerance band over round-level mean scores instead of a fixed hop count or top-k. We evaluate the produced context blocks, not a ranked list, using three LLM judges on a three-dimensional rubric. On a 163-node emotional-companion memory graph with 76 queries, the full system scores 4.46–4.72 against 3.25–3.77 and 2.92–3.46 for two baselines, consistently ranked first by all judges. In an isolated setting that fixes the graph and the assembler and changes only the memory-selection mechanism, the full PAR+story configuration improves over vector top-k by 1.35–1.86 on 70 non-abstention queries. We also report a negative result that constrains the claim: the purpose channel changes the retrieved set substantially, yet its direction is right about half the time, because a purpose represented as a pooled embedding of abstract phrases carries topic but not what kind of content. Rewriting the purpose as three explicit requirements raises the type-distribution divergence between intents from 0.09 to 0.47 and gold-facet recall from 0.30 to 0.53, but those requirements are still supplied by a hand-written mapping. Code, data, prompts, and archived statistics are in the companion software record listed under related identifiers.
Authors
- Siyu Chen
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-19
- DOI
- https://doi.org/10.5281/zenodo.24372784
- Primary Topic
- Information Retrieval and Search Behavior
- Type
- preprint