PROrunner926/copilot-cache-scout

MindMirror — AI Conversation Cache Analyzer

📋 Overview

MindMirror is a diagnostic toolkit that measures the real cognitive cost of multi-turn AI conversations, comparing three distinct caching strategies: naive memory retention, prompt-level caching, and a novel "librarian-pattern digest" that mirrors human memory consolidation. Originally designed as a harness for evaluating token economics during multi-agent code review on Claude Opus 4.6 via the GitHub Copilot proxy, this repository has evolved into a standalone analysis framework for any conversational AI pipeline.

The core insight behind MindMirror is that current token pricing models fail to account for the recursive memory tax—the hidden cost of repeated context injection across turns. By instrumenting the actual token flow through three parallel caching architectures, we reveal where budgets bleed and which patterns conserve the most cognitive bandwidth for downstream agents.

---

🔍 The Three Caching Architectures

🧠 Naive Memory Retention

The baseline. Every turn concatenates the full conversation history without optimization. This is what most chat interfaces do internally. MindMirror measures the exact token overhead as conversation depth increases, exposing the quadratic explosion that occurs beyond 10–15 turns.

⚡ Prompt-Level Caching

The incremental improvement. Only the most recent N turns are retained, with older context summarized by a secondary agent. This mirrors GitHub Copilot's default behavior. Our harness quantifies the token savings versus naive retention, but also surfaces the reconstruction tax—tokens spent re-explaining context that was evicted.

📚 Librarian-Pattern Digest

The novel contribution. I