Can an AI agent remember a correction at the right moment?

A correction such as ‘include the currency next to monetary amounts’ is easy enough to say. The harder part is making sure an assistant applies it the next time it produces a financial answer, even if the company being discussed is different.

Nima Vahdat and Pashootan Vaezipoor investigate this problem for agents that use tools. Their paper presents Forge Decision Context, or forge_dc, a system designed to turn corrections and patterns in tool use into reusable guidance. Rather than changing the underlying language model’s weights, it stores rules separately and supplies them as context when they may be useful. The work is an arXiv preprint accepted at the NeurIPS 2026 Workshop on Personalized, Aligned, Long-Term Memory for AI Systems; its main evidence comes from a benchmark created for the study.

From a correction to a rule

The system looks at two kinds of signal. One is an explicit human correction. The other comes from an analysis of tool calls, such as a failure followed by a successful retry. The authors use a language model to turn correction chains and tool-use patterns into short imperative directives. The system records a trigger for each rule: a point in a past trace where knowing the rule might have prevented the mistake.

These rules are not all simply placed in the initial prompt. The system supplies some at the start of a new task, including formatting rules, and searches again between tool calls as the agent accumulates context. A rule that depends on information returned by a tool can therefore appear after that information becomes available but before the next action. The stored guidance also tracks which actor gave a correction; the system has a procedure for choosing among conflicting rules or abstaining when it cannot resolve them.

This distinction matters because remembering a sentence is not the same as using it appropriately. A rule may need to apply to a different company, retain a condition such as a customer segment, or stay silent in cases where it does not belong.

A controlled test of compliance

The authors built DC-Bench, with 71 scenarios across nine behavioral dimensions and two agent domains: financial analysis and customer support. They recorded agent traces while using mocked tools whose outputs were predetermined. In each scenario, the system received a seed trace with corrections or tool-execution patterns and was tested on another subject. Each scenario began with a fresh memory store.

The test agent was held constant across approaches. A language-model judge scored outputs using scenario-specific weighted rubrics, while programmatic checks enforced some concrete constraints. The researchers ran each scenario–approach pair five times. Their reported score is a compliance measure based on these rubrics, not a percentage of real-world users helped or a guarantee that every action was correct.

In the main experiment, forge_dc achieved 95% average compliance, compared with 55% for the authors’ naive retrieval approach, which retrieves chunks of past traces and supplies them as context. Intermediate versions scored between those results. The paper also reports an average of 159 injected tokens per query for forge_dc versus 1,705 for naive retrieval. That comparison concerns injected context; compiling the rules itself has a cost.

What the result does and does not show

The progression changes several things at once: storage, extraction, rule compilation and delivery. Its final gain cannot by itself isolate how much each individual design choice contributed. The authors also tested additional model backbones and report similarly high compliance in their benchmark, but the scenarios and their tool environment remained controlled.

The benchmark probes whether a correction can transfer to a new subject and whether it activates under the right conditions. It does not establish that the same success rate will hold for arbitrary workplaces, messy live tools or long-running use. The use of an automated judge is another reason to interpret the scores as measurements within this evaluation rather than universal accuracy figures.

The practical idea is nevertheless clear: an agent’s memory may be more useful when it records what to do, under which conditions, and when to surface the instruction—not just what happened in a previous conversation.