Intel / article

Decision Context: Persistent Behavioral Learning for AI Agents

Source: arxiv.org
Author: Nima Vahdat, Pashootan Vaezipoor
Date: Published Nov 9, 2026

Context

Two researchers at Georgian AI Lab built a memory layer for AI agents that stores how to behave rather than what is true, and on their own benchmark it raised agents' average compliance with past corrections to 95%. Today's agents start each session blank, so a user who told an yesterday to paginate an API call or to use a particular revenue metric for enterprise clients has to say it again. Popular memory systems such as Mem0 and Zep mainly recall facts, like names or preferences. The paper argues that storing a correction is easy and that the hard part is knowing when it applies, applying its qualifiers intact, and staying silent when it does not apply.

The system, Forge Decision (forge_dc), turns two kinds of signal into rules. It detects human corrections in conversation logs, using keywords such as "actually" or "instead" with a as fallback, and it has a model read only the agent's tool calls to spot recurring failures or calls that failed and then succeeded on retry. A model then compiles each chain of corrections into one imperative rule, marks it as a formatting or content rule, and records a trigger: the point in the trace where knowing the rule would have prevented the mistake. Formatting rules and rules that match the opening request are injected before the agent starts; other rules are looked up again between tool calls, so a rule that depends on data appears when that data arrives. When different users have given conflicting corrections, the system prefers the current user's rule, otherwise the majority's, and otherwise applies neither.

To test it, the authors built DC-Bench: 71 scenarios across nine behaviours, in a financial-analyst and a customer-support setting, where corrections are given about one company and the agent is then asked about another, with tool outputs fixed in advance. A GPT-4 judge scored each answer against a rubric, averaged over five runs, with Anthropic's Claude Opus 4.7 as the agent. With no memory the agent complied 10% of the time; Mem0 and Zep reached 12% and 15%, two procedural-memory methods adapted from earlier papers reached 34% and 49%, and dumping chunked past logs into retrieval reached 55%. forge_dc reached 95% while injecting about 159 per query against roughly 1,700 for the log-chunk approach. Its mean compliance stayed between 93% and 97% across five other Anthropic and OpenAI models. The authors also report that adding a one-sentence reason to each rule did not help and sometimes hurt, because the model then weighed the rule and occasionally overrode it.

The authors flag the main limits themselves. They built DC-Bench, its grading rubrics, all five of their own systems and the adaptations of the procedural-memory baselines, and they say Mem0 and Zep might score higher with different tuning. Each of their stages adds several components at once, so the paper does not show how much of the gain comes from compiling rules, from triggers or from where the rules are placed. The scenarios cover two domains with mocked tools. Within those bounds, the paper, accepted at a NeurIPS 2026 memory workshop and available on arXiv with its code, gives a concrete recipe for agents that stop repeating a corrected mistake. In this benchmark, short compiled directives beat chunks of past session logs about ten times their size, though which ingredient matters most remains open.

Key takeaways · AI-distilled

The authors argue that existing agent memory systems mainly store facts such as preferences, entities and chat history, not procedural knowledge, so users end up correcting the same mistake at the start of every session.
forge_dc captures signals from human corrections and from the agent's self-assessed tool-call failures, compiles them into imperative RULE directives, and injects each rule at the decision point where it applies.
On DC-Bench, 71 scenarios across nine behavioral dimensions built from real executed agent traces in a mocked-tool environment, forge_dc reached 95% compliance versus 55% for naive retrieval.
Comparing five approaches, the authors identify representation, activation and placement as the central design choices: the strongest systems stored behavior as explicit directives delivered where they apply.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.

Read the source arxiv.org