Editable Multimodal Memory with Reinforcement Learning Management for Long-Horizon Personalized Agents

Abstract

Long-horizon personalized agents must retain heterogeneous evidence without allowing memory overflow, modality fragmentation, stale preferences, or irreversible storage errors. This paper presents an editable multimodal memory governed by an actor-critic manager. Text, visual, and auditory events are normalized into versioned working, episodic, and semantic records. The manager observes budget pressure, relevance, preference consistency, conflict, and consent state, and then filters, compresses, updates, evicts, or retrieves records. A user-facing ledger supports inspection, revision, and selective deletion; tombstones invalidate indexes and block later retrieval. In ten independent streams containing 30,000 interactions and 6,000 queries, the manager achieved 27.7 percent top-one recall and 47.5 percent preference recall. It improved the strongest edit-aware rule by 10.1 and 12.1 percentage points, returned no stale memories, and reduced evictions relative to first-in-first-out storage.