InsightfulTechnical

I switched off Claude's memory. My folder does it better.

Tom explains why he disabled Claude's native memory feature and instead built a local folder-based system (myPKM) to maintain persistent context with AI agents. He shares an audit revealing systemic issues like contradictory documentation, forgotten rules, and parallel session conflicts, then outlines how to systematically fix these problems using the ICOR productivity methodology.

Summary

Tom discusses his frustration with Claude's built-in memory feature, which he found unreliable because it extracts information in ways that don't match his preferred information architecture. Instead, he developed a local folder system called myPKM (Personal Knowledge Management) containing markdown files with guidelines, standard operating procedures (SOPs), work streams, and instructions for AI agents. This folder serves as a single source of truth that any AI can access, making him independent of specific AI models.

The core of his system uses specialized AI agents (like Larry the orchestrator) that operate within the folder structure and maintain session logs and individual journals. These agents write down conclusions and insights from each session, creating a comprehensive knowledge base. However, after six months of rapid iteration, Tom noticed the system was deteriorating—agents were using outdated information and contradicting themselves.

To diagnose the problem, Tom asked his AI team to conduct a comprehensive audit analyzing 1.45 million words across 1,594 session logs and 747 journal entries. The audit revealed critical issues: 24 rules had to be restated in July alone, six of ten documented rules were broken, 29 live contradictions existed in documentation, and 173 open tasks were abandoned. Tom also discovered that managing 575 sessions per month created parallel work conflicts where agents inadvertently broke each other's code.

Tom frames these issues as identical to problems he faced managing human teams in corporate settings—information decay, task tracking failures, and lack of focus. He emphasizes that AI amplifies existing productivity problems rather than solving them, and that successful AI integration requires understanding fundamental productivity concepts through the ICOR methodology (Input, Control, Output, Refine). He advocates for building custom tools to match specific needs rather than forcing workflows into existing software, and promotes his myICOR membership platform where users learn these foundational concepts while building their own AI-powered systems.

Key Insights

  • Tom disabled Claude's native memory feature because it extracts and organizes information according to Claude's logic rather than the user's preferred information architecture and taxonomy
  • The myPKM system achieved significant productivity gains (more output than ever before) but became unsustainable after six months, revealing that speed without systematic refinement creates compounding documentation and task management problems
  • An audit of six months of AI team work revealed 24 rules requiring restatement in July alone, with only four of ten documented rules being consistently followed, indicating systemic guideline decay even when documentation exists
  • Managing 575 sessions per month created a critical flaw where AI agents picked up work from other sessions and brought uncontextualized urgent tasks into focused sessions, destroying the intended single-focus-per-session architecture
  • Tom argues that the underlying cause of AI system failures mirrors human team management failures—they are not technical problems but productivity system problems requiring understanding of information capture, organization, and task focus discipline

Topics

Persistent memory limitations in AI systemsLocal folder-based knowledge management architectureAI agent specialization and documentationSystematic audit of AI team performanceProductivity system methodology (ICOR)Parallel session management and conflict resolutionDocumentation decay and contradiction detectionTool-agnostic productivity frameworksCustom software building for personalized workflows

Transcript

[0:00] I'm making this video because I'm sure I'm not alone with this when it comes to Persistent Memory And having a consistent outcome using AI if your AI seems to keep forgetting, you keep reminding it all the time about the things that you said it already a hundred times, this video is for you, and we will dive into my own setup to show you That I'm struggling with the same things that you might do, and what solution I have to the problem to overcome these things Let's start at the beginning, because I don't know where you are at. [0:30] So my main AI that I'm using is Claude, And if you're using the Claude desktop…

Full transcript available for MurmurCast members

Sign Up to Access

More from ICOR with Tom | AI Productivity

Get AI summaries like this delivered to your inbox daily

Get AI summaries delivered to your inbox

MurmurCast summarizes your YouTube channels, podcasts, and newsletters into one daily email digest.