Unofficial user report: Structured memory system attempt using GPT projects

[Unofficial User Report] Proposal: GPT-Based Information Structuring Experiment

Submitted by: Advanced GPT user (non-developer, general user)
Purpose: To share a real-world usage case with OpenAI or other AI platform developers, offering insight into user-centered AI system design.


1. Objective & Intent

  • This experiment was not aimed at casual use, but to utilize GPT as a personal, rule-based, memory-structured assistant.
  • Without any programming, the user attempted to construct a self-managed knowledge system using only GPT’s project and document features.
  • The final goal was to create something like a “personal AI operating system.”

2. Method Summary

  • Utilized GPT’s project/document functions to create a structured memory system called the “Permanent Storage List.”

  • Applied detailed formatting rules:

    • Hierarchical structure: Numbers → Letters → Korean letters
    • Command-based saving: Nothing is stored unless the user explicitly says “save this”
    • Response constraints: In certain projects, GPT was instructed to refer only to specific documents (e.g., item 3)
    • Controlled formatting: Indentation, spacing, bold emphasis, punctuation rules
  • All stored content was manually confirmed and structured, with no autonomous memory behavior allowed.


3. Discovered Limitations

  • GPT’s document save logic is overwrite-based, meaning new content replaces the entire document unless the model merges previous content manually
  • Deleting and recreating a project causes loss of document linkage, even if document names are reused
  • GPT may modify or delete documents unexpectedly, even with rules that prohibit such behavior
  • GPT does not reliably retain user-defined structural rules across sessions or document interactions

4. Significance of This Attempt

  • This use case likely represents <0.1% of GPT users in terms of structural complexity and intent
  • Treated GPT not as a chatbot but as a personal information infrastructure tool
  • Provides a real-world test of GPT’s limitations for autonomous memory, document control, and agent-like behavior
  • Even as a “failed attempt,” this case clearly exposes system-level constraints and user needs

5. Submission Intent

  • The user has chosen not to continue maintaining this structure, due to system limitations and frustration from data loss
  • However, instead of discarding the effort entirely, it is submitted here as a reference point for future GPT memory/agent development
  • It is hoped that this case can offer valuable insight from the perspective of a power user who operated at the system boundary

This may not be highly impactful, but I hope it contributes in some small way. Thank you for your work.