Under the Hood

This page explains the mechanics behind your knowledge base — how persistence works, how corrections flow, what you can customize, and what is portable.

---

Three Persistence Mechanisms

Claude has no memory between conversations. So how does your KB remember anything? Through three layers of persistence, each with a different purpose:

1. Config (CLAUDE.md) — Always Loaded

Your CLAUDE.md file loads at the start of every conversation. It contains your identity, rules, path registry, and skill references. This is the "who am I and how do I behave" layer.

Persistence: You edit this file manually. It does not change on its own.

Scope: Applies to every conversation.

2. Memory (.claude/memory/) — Corrections and Preferences

The auto-memory system stores things Claude has learned from your corrections. If you tell it "Elaine goes by E in casual conversation" or "don't create action items for FYI updates," those corrections persist here.

Persistence: Claude writes to this automatically when it learns something. You can also edit it manually.

Scope: Applies to every conversation (loaded at startup).

3. KB Files (My_KB/) — On Demand

Your People profiles, meeting notes, decision records, and everything else in the knowledge base. These are loaded only when relevant to the current conversation.

Persistence: Created and updated by skills (process-meeting, file, etc.) or by you directly.

Scope: Loaded on demand — only the files Claude needs for the current task.

How they work together:

Every conversation:
  ┌─────────────┐
  │  CLAUDE.md   │  ← Always loaded. Identity + rules.
  │  (config)    │
  └──────┬──────┘
         │
  ┌──────▼──────┐
  │   Memory     │  ← Always loaded. Corrections + preferences.
  │  (.claude/)  │
  └──────┬──────┘
         │
  ┌──────▼──────┐
  │   KB Files   │  ← Loaded on demand. People, meetings, decisions.
  │  (My_KB/)    │
  └─────────────┘

---

Two Correction Paths

When something in your KB is wrong, there are two ways to fix it — and they have different effects.

Path 1: Direct Edit

Open the file, change the text, save.

## Before
Elaine is concerned about the Q2 deadline.

## After
Elaine is concerned about the Q3 deadline.

What happens: The file is corrected. That is it. Claude will read the corrected version next time it loads the file. No learning occurs — Claude does not know the old version was wrong.

Best for: Factual errors, typos, wrong dates, misattributed quotes.

Path 2: Comment Marker

Add an inline comment using your configured marker (e.g., >Jerry:), then ask Claude to process comments.

Elaine is concerned about the Q2 deadline.
>Jerry: This was Q3, not Q2. Elaine specifically said Q3 in the meeting.

When you run "process comments," Claude:

1. Reads the comment
2. Makes the correction in the file
3. Learns from the correction (adds to memory)
4. Removes the comment marker

What happens: The file is corrected AND Claude learns. Next time it processes a similar meeting, it is less likely to make the same mistake.

Best for: Recurring errors, pattern corrections, teaching Claude your preferences.

When to use which:

Situation	Use Direct Edit	Use Comment Marker
Simple typo	X
Wrong date	X
Misattributed statement		X
Claude keeps making same error		X
You want Claude to learn a preference		X
Quick fix during a review	X

---

The Config File: CLAUDE.md

CLAUDE.md is the most important file in your setup. Here is what goes in it:

Profile Section

## Profile
- Name: Jerry
- Role: Engineering Lead
- Team: Vandelay Industries
- Focus: Import-Export project

Path Registry

Tells Claude where everything lives:

## Paths
| Key | Path |
|-----|------|
| PEOPLE | My_KB/People/ |
| MEETINGS | My_KB/Meetings/ |
| DECISIONS | My_KB/Decisions/ |
| WORKBOARD | My_KB/Work_Board.md |
| INBOX | My_KB/Inbox/ |
| PROJECTS | My_KB/Projects/ |
| AREAS | My_KB/Areas/ |
| RESOURCES | My_KB/Resources/ |
| ARCHIVE | My_KB/Archive/ |
| TEMPLATES | My_KB/Templates/ |

This registry is critical. When a skill says "save to MEETINGS," Claude looks up this table to find the actual path. If you reorganize your folders, you update the registry and everything keeps working.

Extraction Rules

## Extraction Rules
When processing meetings, extract:
- Positions expressed by each speaker
- Concerns and risks raised
- Decisions made (explicit and implicit)
- Action items with owners and deadlines
- Patterns across recent meetings

Do not extract:
- Off-the-record discussions
- Personal salary or compensation details
- Unverified rumors or speculation

Response Protocol

## Response Protocol
- Be direct and concise
- Cite sources (link to files) when making claims
- Ask before overwriting existing content
- Flag uncertainty rather than guessing

Skill References

## Skills
See .claude/skills/ for available skills.
Core skills: process-meeting, file, note, start-day, end-day, week-review

---

Where Things Are Stored

A complete map of where everything lives:

What	Where	Purpose
Config	CLAUDE.md	Identity, rules, path registry
Skills	.claude/skills/	Reusable instruction sets
Memory	.claude/memory/	Auto-learned corrections and preferences
KB files	My_KB/	People, meetings, decisions, projects
Templates	My_KB/Templates/	Blank templates for new files
Settings	.claude/settings.json	Claude Code harness configuration

---

What You Can Customize

Different parts of the system have different levels of customizability:

Safe to Change (Go Ahead)

These are meant to be customized. Change them freely.

• CLAUDE.md content — your identity, rules, extraction preferences
• Skill files — modify existing skills or create new ones
• Templates — change the format of meeting notes, People profiles, etc.
• Folder structure — reorganize My_KB/ however you like (update path registry)
• Frontmatter schemas — add custom fields to any file type
• Work Board format — change the Kanban columns, add sections
• Extraction rules — control what gets extracted and what does not

Change Carefully (Test First)

These work but need some care.

• Path registry — if you change paths, make sure skills still reference the right locations
• Skill triggers — changing trigger words can affect when skills fire
• Memory entries — editing auto-memory can change Claude's learned behaviors
• Settings.json — harness-level config affects tool behavior

Managed by the System

These are maintained automatically. Edit only if you know what you are doing.

• Auto-memory index — managed by Claude's learning system
• Skill header caching — managed by the harness
• Tool permissions — managed by MCP configuration

---

What Is Portable (Open Standards)

One of the most important questions: what happens if you want to switch tools?

Fully Portable (Open Standards)

These use open formats and work anywhere:

• KB files — Plain Markdown with YAML frontmatter. Readable by any text editor, Obsidian, Notion (via import), and any future AI system.
• Skills — Plain Markdown files. The instruction format is generic enough to adapt to any LLM.
• MCPs — Model Context Protocol is an open standard. MCP servers work with any compatible client.
• Frontmatter — Standard YAML. Parseable by any programming language.
• Wiki-links — Supported by Obsidian, many Markdown editors, and easily parseable.
• Folder structure — Just folders and files. No database, no proprietary format.

Partially Portable

These work with the current system but would need adaptation:

• CLAUDE.md — The file is plain Markdown, but its specific format (path registry, extraction rules) is designed for Claude Code. Another AI system would need similar configuration in its own format.
• Auto-memory — The memory system is specific to Claude Code. The contents are readable (plain text), but the automatic loading mechanism is tied to the harness.
• Skill triggers — The slash-command and natural-language matching is specific to Claude Code. The skill instructions themselves are portable; the trigger mechanism is not.

The Bottom Line

Your data is yours. Everything important lives in plain text files on your computer. If you want to switch from Claude to another LLM, you bring your KB files, rewrite your config for the new system, and adapt your skills. The knowledge — the actual content — transfers with zero friction.

This is by design. A personal knowledge base should outlast any single tool.