Building a Memory System

Table of content

AI without memory is just a fancy calculator. A personal OS needs to remember.

Memory types

Type	What it stores	Example query
Episodic	Specific conversations, events	“What did we discuss last Tuesday?”
Semantic	Facts about you, preferences	“What time zone am I in?”
Procedural	How to do things	“When I say ‘deploy’, run these commands”

Option 1: Episodic Memory Plugin (Recommended)

claude mcp add episodic-memory
claude mcp status  # Verify

# Test
claude "Remember that I prefer dark mode"
claude "What are my preferences?"

Data stored locally in ~/.claude/memory/

Option 2: Memory file

Create ~/.claude/memory.md:

## About me
- Software developer, Europe
- Prefer async communication

## Decisions
- 2025-01-15: Chose React over Vue (team familiarity)
- 2025-01-10: Work hours 9-18

## Active projects
- ProjectX: React dashboard (deadline: Feb 2025)

Reference in CLAUDE.md:

Always read ~/.claude/memory.md for context.
Update when I share preferences or decisions.

Option 3: External tools

Tool	Install	Best for
Obsidian	`claude mcp add obsidian`	Existing Obsidian users
Notion	`claude mcp add notion`	Structured databases

Memory hygiene

Action	Command
Review outdated info	`claude "Review memory and suggest cleanup"`
Save important decision	`claude "Save to memory: [decision]"`
Query memory	`claude "What do you remember about X?"`
Fix wrong info	`claude "Update memory: correct info is..."`

Building context over time

Timeline	What happens
Week 1	Basic preferences saved
Week 4	Patterns emerge
Month 3	AI understands your work style
Month 6	Feels like a knowledgeable assistant

Privacy

Episodic memory plugin stores data locally
Cloud sync (Notion/Obsidian) sends data to their servers
API calls send content to Anthropic

For sensitive data: use local-only storage, don’t store credentials.

Next: Custom Commands

Topics: memory claude-code mcp