AI agents need more than facts — they need institutional memory capturing why decisions were made. Using three variants of the classic Zebra Puzzle as a framework, the post explores three layers of agent reasoning: strict logical constraints (rules engines), natural language ambiguity requiring probabilistic reinterpretation, and subjective/affective constraints requiring human-like judgment. Graphs serve as a unified structure for knowledge, computation, and judgment simultaneously. The post argues that context graphs must capture not just what decision was made but the subjective reasoning behind exceptions, especially for enterprise agents handling mixed hard rules, fuzzy guidance, and unresolvable tensions. Research combining LLMs with formal constraint solvers showed a 166% improvement in puzzle-solving for GPT-4.
Table of contents
Where does the Zebra live?Get Andreas Kollegger’s stories in your inboxWhy Graphs? Knowledge and Thinking in One StructureWhat’s coming nextWhere This Goes NextSort: