Remembering What Matters: Gated and Bounded Memory for Agentic Systems

Large language model (LLM) agents are increasingly deployed in long running workflows, where they must preserve user and task state across many turns. Many existing agent memory systems treat memory primarily as a storage and retrieval problem, leaving less emphasis on deciding what should be remembered. We develop a gated and bounded memory architecture that makes memory formation selective. An LLM based gate filters incoming information based on its relevance to the agent's current goal and assigns utility, while a bounded episodic buffer preserves recent interactions. High utility experiences are periodically consolidated into a structured knowledge graph, while low utility traces are discarded. We evaluate the architecture under clean inputs and injected distractor noise and find that it achieves higher noisy F1 and smaller performance drops than Vanilla RAG and Mem0, suggesting that selective encoding, bounded storage, and consolidation can improve the robustness of agent memory.