Claude Memory Skill

claude-memory-skill
*An embarrassingly simple and minimal implementation for agentic memory.*
No databases. No embeddings. No semantic search. Just markdown files and a skill that teaches Claude when to read and write.

The Problem

Claude Code forgets everything between sessions. Built-in auto-memory exists but:
- It's opaque (Claude decides what's "meaningful")
- Limited to 200 lines loaded at startup
- Not tightly integrated into the agentic loop
- No hierarchical organization (scales poorly)

The Solution

A skill-based memory protocol with:
- Hierarchical storage: core.md summaries → topics/.md details
- Background agents: Memory ops don't block the main agent
- Categorized entries: No dumping ground, everything has a topic
- Filesystem-based: Robust, inspectable, git-trackable

Installation

Or clone and run locally:

What Gets Installed

Architecture

How It Works

Agent-Initiated (Automatic)

These run automatically — you don't invoke them:

User-Initiated (Manual)

These are for you to inspect and manage memory:

Tell Claude What to Remember

Just say it naturally:
- "Remember that we use pnpm, not npm"
- "Save that the API requires auth headers"
- "Note that tests need Redis running"

Fill In Your Profile

Edit ~/.claude/memory/me.md with facts about yourself:

Hierarchical Memory

core.md (always loaded):

topics/debugging.md (loaded on demand):

Write Flow

1. Learn something → spawn background memory agent
2. Agent categorizes → finds or creates topic file
3. Agent appends entry with timestamp
4. If significant, agent updates core.md summary

Read Flow

1. Session start → agent reads core.md + me.md in background
2. When stuck → agent follows pointers to relevant topics
3. Returns context to main agent

Design Principles

- Background ops: Memory doesn't…