Will Lyon separates short-term conversation memory, long-term knowledge and reasoning memory that records the evidence, policies and tool results behind a decision. Will Lyon argues that entity resolution and a shared ontology give agents a connected view of these records, rather than relying only on similarity search over disconnected text.
Will Lyon describes a pipeline from recorded experience to candidate skills, then to grounded steps and reusable execution structures. Markdown skills offer progressive disclosure, but typed graphs can make dependencies and execution paths easier to inspect. Research results discussed in the talk are presented as supporting examples, not proof that every automatically distilled skill will outperform a human-written one.
Will Lyon's healthcare example turns patient-intake conversations into a charting workflow with explicit tool and entity references. The broader engineering lesson is to evaluate grounding, coverage and coherence, restrict memory to the right scope, and detect when a skill's underlying data or tools become stale. The demonstration concerns agent architecture, not clinical advice.
Watch on YouTube
