2025-07-24 AI Mastermind

Table of Contents

Session Overview

This session opened with Kasimir’s early-stage experiment in persistent AI memory: a local MCP database that captures high-impact learnings from each Claude session, stores writing style guidelines, and implements a voice authentication check to ensure output meets a threshold similarity to his personal voice. The system had only been running for a couple of days, but the architecture was sophisticated enough to generate a rich discussion about what cross-session AI intelligence could look like over six months of use.

Don Back then shared his fully-developed content production workflow: from psychographic avatar through content pillars through topic calendar through multi-format draft generation through physical pencil editing. This workflow is notable for its clarity about where AI adds leverage versus where human judgment is irreplaceable — and for the evidence that it is producing content that resonates with Don’s LinkedIn audience.

Lou extended both threads: he introduced Mem0 as an established implementation of the memory concept, and offered the “story library in memory” idea as a way to automate the hook-story-offer content structure. He also coached Don on the voice profile training approach — using before/after editing pairs to train the model on his personal style, eliminating most of the revision work.

A significant sub-discussion explored the MCP ecosystem: ChatGPT announced MCP support in mid-2025, which means Kasimir’s memory architecture can now be portable across all major AI platforms. The vision of a single source of truth for your personal AI context — your stories, your voice, your knowledge, your brand — accessible across every tool, is now technically feasible.

High-Signal Moments

  • Kasimir’s voice authentication threshold: output must be 70-80% similar to his writing style or it is automatically flagged and revised
  • The “threes and sixes” pattern — AI defaults to lists of 3 or 6 items; Kasimir’s system now flags this as an AI telltale and triggers alternative structures
  • After 20 messages: automatic lesson capture. After 3 sessions: pattern analysis. This is learning systematized, not left to chance
  • ChatGPT now supports MCP (mid-2025 announcement) — making cross-platform memory portable for the first time
  • Don’s content production workflow: psychographic anchor → 10 content pillars → 6-month topic calendar → 3-format draft → pencil edit → publish
  • The three-format draft (conservative / middle / edgy) as a mechanism for preventing default-to-safe content choices
  • Lou’s “story library” concept: store your personal stories in memory, have the AI retrieve the most relevant one when writing content to anchor the hook-story-offer structure
  • Lou’s voice profile training technique: 10-12 before/after editing pairs → let AI derive the style guide → dramatically reduces revision time
  • Don reporting genuine LinkedIn engagement improvement from AI-assisted content — the human editing step matters
  • Hook-story-offer (Tony/Dean structure) referenced as the content framework most worth systematizing

Open Questions

  • What is the realistic timeline for a self-improving memory system to produce meaningfully differentiated output from a generic AI?
  • How do you handle the risk that a voice authentication system becomes a cage rather than a guide — preventing the AI from writing in ways that stretch your voice productively?
  • What’s the right level of granularity for a story library — 5 foundational stories or 50 situational ones?
  • As MCP becomes the standard for AI tool integration, what does this mean for data portability and privacy across platforms?
  • What should a coach’s minimum viable personal AI memory architecture look like to get started today?

Suggested Follow-Through

  • Explore Mem0 (mem0.ai) as a starting point for a local memory architecture
  • Create your Layer 1 Story Library: document 5-10 foundational personal stories, tagged by lesson/theme
  • Begin your voice profile: identify 5 pieces of writing that best represent your ideal voice; identify 5 examples of AI output you’d want to edit away from
  • If you have before/after content editing pairs from past AI-assisted writing, gather 10-12 of them for voice profile training
  • Test ChatGPT’s MCP support — explore whether your existing Claude MCP tools are accessible there

Additional Resources

Ideas from Chat

  • Donald Kihenja noted that giving Claude a “big story” lets it weave relevant fragments into many pieces of content — a practical extension of the story library concept
  • Donald’s idea of using Firebase Studio to create, deploy, and get a free URL in one flow — a zero-cost deployment path for small tools
  • “Hallucination layering” — the risk that compounding AI outputs amplify initial errors (noted by Bally Binning as a concern worth tracking)