Abstract
Systems and methods are described for managing natural-language preference memory in an agentic recommendation system. A preference memory stores per-user preference chunks as natural-language statements annotated with metadata including confidence with temporal decay, rolling-window exposure count, engagement-quality trend, a source tag distinguishing explicit versus inferred preferences, and a derived satiation indicator. An effective preference weight is computed from decayed confidence, a source-dependent multiplier, and a satiation-based discount, and is supplied to a large language model (LLM) ranking agent to modulate candidate scoring. Preference lifecycle states may be assigned based on effective weight thresholds, and profile summaries may be generated from weighted active preferences while excluding saturated and retired entries. The preference chunks may further include a challenge date and challenge history enabling scheduled boundary-testing content with adaptive geometric backoff based on user responses.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Anonymous, "Metadata-Annotated Preference Memory Architecture for Recommendation Systems", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/10750