Enterprise AI agents keep failing because they forget what they learned

Enterprise AI agents keep failing because they forget what they learned">

RAG architectures have a singular strength: surfacing semantically relevant documents. However, their capabilities stop short of providing the decision context necessary for agents to make informed choices. This limitation stems from their inability to offer structured memory, time-aware reasoning, and explicit decision logic.

Rippletide, a startup within the Neo4j ecosystem, has developed a framework that addresses these shortcomings. Their approach centers on a decision context graph, which equips agents with non-regressive capabilities, allowing them to build on validated sequences of actions over time. According to Yann Bilien, Rippletide's co-founder and chief scientific officer, the key is non-regressivity.

"How do you make sure that, when the agent will generate something new, you can compound on the previous discoveries?" he asked. The challenge with RAG architectures lies in their limited scope. Enterprise context is dispersed across various tools, logs, databases, vector stores, and policy documents.

While generative AI tools can retrieve information from these sources, they often struggle with relevance and guidance. Wyatt Mayham of Northwest AI Consulting noted that RAG architectures "retrieve documents, not decision context." This oversight leads to agents combining incompatible rules, inventing constraints, and relying on probabilistic guesses. Decision context graphs offer a more robust solution by encoding a structured map of applicable rules, their validity, and temporal context.

This framework is optimized to answer: "Given this situation, which context applies right now?" Time is treated as a critical dimension, with every rule, decision, and exception scoped to its validity period. Bilien emphasized that the goal is to explicitly address missing, incoherent, or contradictory data when building the graph, thereby avoiding probabilistic errors. The decision context graph system operates on three core principles: applicability, time-aware memory, and decision paths.

It provides agents with the ability to reason about past and present contexts, reproduce decisions, and explain their rationale. Through neuro-symbolic AI, the system recognizes patterns and encodes formal logic. As new decisions are made, the system refines its knowledge base.

The benefits of this approach include non-regressive learning, where agents can explore, validate behaviors, and compound on intelligence and knowledge. This leads to more consistent, predictable, and explainable behavior, crucial for reliability at scale. According to Bilien, "You want your agents to be able to learn by themselves when they face something they don't know." The system ensures that agents can generalize solutions across similar cases while preserving previous capabilities.