An agent can retain a large history without being able to use the distinction that matters for its next task. The Value of Memory in Language Agents develops a decision-theoretic framework that follows memory from raw experience through persistent state, operational use and costly recovery or revision.
Measure value in the task
The paper expresses original and retained value in units of avoided task loss. It uses risk-cost frontiers to describe tradeoffs in preserving information and recovering it later. Unknown future demand is represented through value profiles, while interactive feedback can change the distribution of later work.
The point is that a memory’s size or compression ratio alone does not establish its usefulness. An evaluation needs to specify what future decisions require and what it costs the bounded agent to retrieve or reconstruct the relevant evidence.
A pointer can matter without containing the answer
The authors use a cost-aware simulation condition and five minimal counterexamples to examine timing, retrieval access, component interactions, feedback and uncertain demand. One example shows that a pointer with zero mutual information about the answer can still reduce the cost of exact recovery.
That separates storing an answer from preserving a useful route back to it. In a real workflow, a source reference may be valuable because it makes correction or verification possible, even when the saved summary is deliberately brief.
Keep memory effects separate from scoring effects
Diagnostics include LoCoMo, a two-project update task and longitudinal cases. The paper reports that a saved ledger can change subsequent answers at fixed retrieved evidence, while an evaluator revision can change scores at fixed agent outputs.
These are different effects. A higher score does not by itself prove that the memory improved the underlying task result. A useful evaluation would preserve the evidence retrieved, the state carried forward, the agent’s output and the evaluator version separately. The framework offers ways to reason about those distinctions, rather than identifying one memory architecture as best for every agent.
Comments