hanfang/claude-memory-skill
FreeHierarchical memory system with filesystem persistence
FreeFree tier
About hanfang/claude-memory-skill
An embarrassingly simple and minimal implementation for agentic memory for Claude Code. No databases, no embeddings, no semantic search. Just markdown files and a skill that teaches Claude when to read and write. Features hierarchical storage with core.md summaries and topic directories, background agents for non-blocking memory operations, and filesystem-based persistence that is robust, inspectable, and git-trackable.
Key Features
Hierarchical storage: core.md summaries point to detailed topic files
Background memory agents: memory operations don't block the main agent
Categorized entries: everything organized by topic, not a dumping ground
Filesystem-based: robust, inspectable, and git-trackable
Automatic memory operations: load on session start, save on learning, recall when needed
Manual memory commands: /mem show, /mem forget for inspection and management
User profile file (me.md) for personal context
Project-specific memory directories
Simple bash-based installation script
Pros & Cons
Pros
- No databases, embeddings, or semantic search - extremely simple
- Non-blocking background memory agents
- Hierarchical organization scales better than flat memory
- All memory is plain markdown - fully inspectable and editable
- Git-trackable for version control of memory
- Easy one-command installation
Cons
- Only works with Claude Code, not a general-purpose memory system
- Manual commands required for inspecting and managing memory
- Relies on Claude's ability to follow the skill protocol correctly
- Not suitable for very large-scale or high-frequency memory operations
Best For
Persistent memory for Claude Code across sessionsKeeping project-specific context and notesMaintaining knowledge of user preferences and workflowDebugging history and solutions trackingResearch notes and experiment tracking
FAQ
What is claude-memory-skill?
A minimal, low-friction hierarchical memory system for Claude Code that uses markdown files instead of databases or embeddings.
How does it work?
It installs a skill that teaches Claude when to read and write memory files. Memory operations are handled by background agents so the main agent is not blocked.
What gets installed?
The installation sets up ~/.claude/ with a CLAUDE.md hook, a commands/mem.md skill, and a memory/ directory containing core.md, me.md, topics/, and projects/.
Can I manually manage memory?
Yes, you can use commands like /mem show to display memory structure and contents, and /mem forget to remove a topic or specific entries.
How do I tell Claude to remember something?
Just say it naturally, for example: 'Remember that we use pnpm, not npm' or 'Save that the API requires auth headers'.