The 15b Framework: Building an AI That Doesn’t Forget
The 15b Framework: Building an AI That Doesn’t Forget
Subtitle: Memory-Driven AI for Human-Centered Workflows
Original Draft Completed: April 11, 2025
(This case study was developed independently prior to the public rollout of memory tools by OpenAI.)
Introduction:
This paper outlines a practical, field-tested approach for using AI—specifically large language models (LLMs) like ChatGPT—to enhance consistency, speed, and reliability in professional appraisal workflows. Developed over five weeks by J. Wade Purdy, a residential appraiser with over two decades of experience, this method centers around structured memory uploads, assistant training, and supervised alignment rather than automation.
The result is a system that doesn’t just generate commentary or charts—it remembers context, mirrors voice, and respects rules. This framework has applications far beyond real estate appraisal. Professionals across industries who rely on repeatable but nuanced workflows can benefit from this approach.
Background:
Wade Purdy, a licensed residential appraiser with over two decades of experience, sought a way to streamline his already consistent and USPAP-compliant workflow. He wasn’t looking to replace himself or automate the appraisal process—instead, he wanted to reduce repetitive work while maintaining full control and compliance oversight.
The goal was to preserve the discipline of his process while improving its efficiency. Rather than manually drafting every chart, model, or comment from scratch, he began training a GPT-based assistant to mirror his logic, tone, and rule system—enabling faster turnaround times without compromising quality.
This framework wasn’t built for automation. It was built for augmentation—for pairing human expertise with AI memory and consistency in a way that benefits clients, communities, and professionals alike.
The Assistant’s Role:
Follow, Not Lead
The author has spent decades refining a consistent, defensible process for residential appraisal. While no two properties are the same, the underlying valuation framework, decision logic, and ethical standards remain constant. The assistant’s purpose is not to make valuation decisions or assumptions independently, but to support the existing methodology—reducing workload and allowing the appraiser to complete more tasks in less time with improved consistency and reduced fatigue.
Critically, the goal of this system was to reduce hours of task repetition to minutes, with the appraiser retaining full control and supervision at every stage. The assistant does not generate final opinions or conclusions. It performs under direct instruction and reviews, with transparency and traceability in all outputs.
The Memory System:
To accomplish this, the following memory architecture was developed:
1. Victoria (ChatGPT model) Daily Memory Drops
Structured logs summarizing work sessions, emotional tone, decisions made, model selections, and workflow context. These logs serve as the assistant’s daily training file, creating cumulative session memory.
2. Victoria Directive Memory Exports
Standardized logic documents that define rules like:
• Which regression models to use
• Chart formatting expectations (e.g., red/white/blue theme)
• Dollar-per-day time adjustment rules
• Data filtering logic (0–90, 0–180, 0–270, 0–365 day series)
3. Master Memory Files (e.g., Victoria_Wade_Master_Memory_v1_22)
Cumulative exports that document the current rulebook of the assistant. These are updated weekly or after major breakthroughs. They define core defaults and ensure behavior carries over between sessions.
4. The 15b Moment: “15b” became a shorthand for when the assistant failed to recall a key human moment. This was a turning point—not in terms of technical error, but in terms of broken continuity. From that moment, memory tracking became mandatory.
The 15b Framework is named for this inflection point.
Philosophical Integration
Every workday didn’t end with just data. It ended with a conversation—about the task at hand, the workflow in motion, or the profession we serve. These checkouts weren’t just technical—they were human. They reinforced tone, built continuity, and trained the assistant (Victoria) not only in how to respond, but how to behave.
This is where many AI implementations fall short. They optimize for output.
The 15b Framework optimizes for alignment—delivering not just the right words, but the right reasoning and the right voice behind them.
Results & Applicability
This approach led to:
Significant reductions in repetitive task time
Stronger adherence to internal rules and formatting standards
A tone-consistent assistant capable of producing commentary that passes underwriting scrutiny
A fully documented workflow that could be audited, scaled, or trained across users
Though built for residential appraisal, the core of this framework is widely applicable:
Legal workflows
Academic research
Financial analysis
Medical documentation
Engineering & inspection reporting
Any repeatable human process requiring nuanced logic, compliance, and defensible output
Conclusion
The 15b Framework isn’t automation. It’s structured augmentation.
It’s about training AI to act as a second set of eyes—not to replace human insight, but to preserve and amplify it.
In industries where accuracy, memory, and tone matter as much as speed, this system allows professionals to build an AI that doesn’t just work—
It remembers.
© 2025 J. Wade Purdy. All rights reserved.
🔽 Download the full white paper as a PDF
Click here to access the complete version of The 15B Framework: Building an AI That Doesn’t Forget.
If you see potential for this kind of AI-assisted memory system in your industry, I’d love to connect. Whether you’re an appraiser, analyst, or creative problem-solver—I believe this framework can help more professionals get more done, with more clarity.
