Lecture 05: Multi-Session State Persistence

Why Agents Suffer from Amnesia

In real software development, complex technical SEO sprints cannot be solved in a single prompt or single session. They require:

• Phase 1: Baseline Audit & Safety Deployment.
• Phase 2: Schema.org Graph Architecture.
• Phase 3: Robots & Canonical Normalization.
• Phase 4: Core Web Vitals Optimization.
• Phase 5: Verification & Testing.

When you start a new conversation with an agent, its context window is empty. It has zero memory of what was done previously.

---

The Dual-File State Pattern

The Genisoft Harness solves this through two files persisted to the local filesystem:

1. seo_tasks.json (Machine-Readable Backlog)

{
  "id": "SEO-02",
  "phase": "Schema.org Architecture",
  "title": "Implement Connected Knowledge Graph (@graph)",
  "status": "completed",
  "verification_gate": "node tools/schema-graph-validator.js build/schema-home.json",
  "evidence": "3/3 entities validated with 0 dangling @id references."
}

2. progress.md (Human/Agent-Readable Session Log)

• Records what was changed in each session.
• Records files modified and git commit hashes.
• Designates the exact task ID to be picked up in the next session.