Forget to Improve: On-Device LLM-Agent Continual Learning via Budget-Curated Memory

On-device LLM agents improve by accumulating experience in memory rather than by updating model weights. However, this memory is constrained by limited RAM, energy, and bandwidth, and it is vulnerable to poisoning attacks. The authors ask: can a single, unified governance mechanism manage memory retention, cross-agent sharing, and trust to optimize performance under these strict edge-computing constraints?

The authors propose CURATOR, a framework that evaluates every memory entry using a single metric: net-value-per-byte. This score calculates the expected future value of an entry minus a harm term (which accounts for negative transfer and provenance risk), normalized by the entry's byte size. This single score governs three critical decisions:

• KEEP: Evicts low-value entries to stay within RAM and energy budgets.
• SHARE: Selects only high-value insights to transmit over limited network uplinks.
• TRUST: Gates incoming peer data by checking provenance to prevent injection attacks.

The system was evaluated on LLM-agent task-drift benchmarks and a real-world heterogeneous Jetson testbed consisting of robot-arm nodes and a central hub.

CURATOR demonstrates that forgetting is not necessarily detrimental to an agent's capability; rather, it is a form of active editing that improves performance. Key results include:

• : Reduced peak memory usage by 2.7x and uplink bandwidth by 2.4x compared to non-curated baselines.