Persistence Patterns

Verify-then-Persist

A step that emits structured JSON, gated by a jsonMatches schema, with a write action on the gate. Turns a stateless ladder into a learning system.

Shape and when to use

Shape: a step emits structured JSON, an if: { jsonMatches: ... } gate verifies the JSON matches a schema, and a then: { write: ... } action persists the verified output to a slate.

When to use: durable learning across calls. Record what worked, what didn't, structured outcomes that future invocations can read via slateRead or retrieve:.

Cost: 1 LLM call (no extra cost; the gate and write run after the step). The gate evaluation and slate write are sub-millisecond operations against the hot tier.

Config sketch

yaml

- id: extract
  type: normal
  fields:
    - { name: Context, type: text, from: input.context }
  systemPrompt: |
    Output JSON summarizing the key finding:
    {"fact": "...", "confidence": 0.0-1.0}
  if:
    jsonMatches:
      type: object
      required: [fact, confidence]
      properties:
        fact: { type: string, minLength: 1 }
        confidence: { type: number, minimum: 0.6 }
      additionalProperties: false
    then:
      write:
        to: { slate: "Memory", folder: facts, file: core.md }
        from: output
        on: append
    else: continue

The gate-and-write pattern is what turns a stateless ladder into a learning system. Every call contributes durable memory, but only for outputs that pass verification. Bad outputs (low confidence, missing fields, malformed JSON) are silently dropped via the else: continue path.