Wire the cache into a session

import FoundationModels

// Illustrative session shape. Verify names in the installed SDK.
@MainActor
final class CachedAsk {
    let embedder: QuestionEmbedder
    let session: LanguageModelSession
    var rows: [CacheRow] = []
    var stats = CacheStats()
    var minimum: Float = 0.92

    func answer(_ question: String) async throws -> String {
        let vector = try await embedder.vector(for: question)
        stats.lookups += 1
        if let row = lookup(question: vector, rows: rows, minimum: minimum, now: .now) {
            stats.hits += 1
            return row.answer
        }
        stats.misses += 1
        let response = try await session.respond(to: question)
        let text = String(describing: response)
        rows.append(CacheRow(
            id: UUID(),
            question: question,
            answer: text,
            embedding: vector,
            createdAt: .now,
            expiresAt: .now.addingTimeInterval(30 * 60)
        ))
        return text
    }
}

If the session uses Dynamic Profiles, the cache still sits outside the profile body. Profiles pick instructions, tools, and the model. The cache decides whether to call the session at all. One Profile is active when you do generate. See Run an agent with a state machine for that rule.

Do not cache tool-using turns until you can key the row on tool results as well. A cached "next meeting" that ignored a calendar refresh is a product bug.

Sources: WWDC26 session 241 (usage and models), Foundation Models, and Core AI.