Remembering Has a Price: Learning Memory Worth under Resource Constraints
TokenWorth achieved the highest task success rate while using the fewest output tokens among the memory methods tested.
The authors evaluated it across three LifelongAgentBench environments, using both local and API-based deployments. Under tight generation, interaction and memory budgets, several existing approaches offered little advantage over simple retrieval or agents without memory. A relevant but ineffective memory can waste scarce context and hinder execution. TokenWorth learns which memories help from sparse execution feedback, then uses those estimates to guide retrieval and decide what to retain.
HF Daily Papers' note
The authors evaluated it across three LifelongAgentBench environments, using both local and API-based deployments. Under tight generation, interaction and memory budgets, several existing approaches offered little advantage over simple retrieval or agents without memory. A relevant but ineffective memory can waste scarce context and hinder execution. TokenWorth learns which memories help from sparse execution feedback, then uses those estimates to guide retrieval and decide what to retain.
HF Daily Papers' note
score 4