Agent Memory Skill

A persistent memory system for AI agents, implementing cognitive memory architecture with semantic, episodic, and procedural memory types. Designed for use with Claude Code.

Overview

LLMs have vast but frozen knowledge—they cannot learn by updating weights after training. Without memory, an agent is like "an intern with amnesia": brilliant but unable to recall previous conversations or learn from experience.

This skill implements a 4-layer cognitive memory architecture:

Internal Knowledge (LLM weights) ← frozen
         ↑
Context Window (per inference) ← this skill helps populate
         ↑
Short-Term Memory (session state)
         ↑
Long-Term Memory (persistent storage)
├── Semantic (facts)
├── Episodic (experiences)
└── Procedural (workflows)

Memory Types

Type	Purpose	Example
Semantic	Facts, preferences, relationships	`"User prefers dark mode"`
Episodic	Timestamped experiences	`"[2025-01-15] Debugged JWT expiry issue"`
Procedural	Workflows, multi-step procedures	`"Deploy: 1. Test, 2. Build, 3. Push"`

Installation

Prerequisites

Python 3.10+ - Required for the memory system
pipx - Recommended for isolated package installation

# Install pipx if not already installed
python3 -m pip install --user pipx
python3 -m pipx ensurepath

# Verify installation
pipx --version

Option 1: MCP Server (Recommended)

This exposes memory operations as tools directly in Claude Code.

Step 1: Install dependencies

# Install MCP SDK (required)
pipx install mcp

# Install CogDB for graph backend (recommended)
pipx install cogdb

Step 2: Clone or download the skill

git clone https://github.com/btafoya/claude-code-skills.git
cd claude-code-skills/agent-memory-skill

Step 3: Configure Claude Code

Add to your Claude Code MCP config (~/.config/claude-code/config.json):

{
  "mcpServers": {
    "memory": {
      "command": "python",
      "args": ["/path/to/agent-memory-skill/scripts/mcp_server.py"]
    }
  }
}

Available MCP Tools:

memory_add_fact - Store semantic memories (facts about user/project)
memory_add_episode - Record significant interactions with timestamps
memory_add_procedure - Save reusable workflows with steps
memory_search - Search memories by keyword
memory_get_context - Build context from relevant memories
memory_stats - Get memory statistics
memory_list_all - List all stored memories
memory_delete - Remove a memory by ID

Step 4: Restart Claude Code

After updating the config, restart Claude Code to load the MCP server.

Step 5: Verify installation

In Claude Code, run:

memory_stats

You should see output like: Total: 0 | Semantic: 0 | Episodic: 0 | Procedural: 0 | Backend: graph

Option 2: Python Library

For direct Python integration in your own projects:

Step 1: Install dependencies

# Graph backend (recommended)
pipx install cogdb

# Or for zero-dependency fallback, skip cogdb

Step 2: Add to your project

# Copy the scripts directory to your project
cp -r agent-memory-skill/scripts your-project/

# Or add to Python path
export PYTHONPATH="${PYTHONPATH}:/path/to/agent-memory-skill"

Step 3: Use in your code

from scripts.memory import MemoryStore

store = MemoryStore()  # Uses graph backend by default

# Add memories
store.add_fact("User prefers TypeScript over JavaScript")
store.add_episode("Helped debug authentication issue", topic="auth")
store.add_procedure("deploy", ["Run tests", "Build", "Push", "Deploy"])

# Search and retrieve
results = store.search("TypeScript")
context = store.build_context("authentication")  # For prompt injection

# Use JSON backend if CogDB not installed
store_json = MemoryStore(backend="json")

Option 3: CLAUDE.md Injection

Export memories for manual inclusion in your project's CLAUDE.md:

cd /path/to/agent-memory-skill
python -c "from scripts.memory import get_store; print(get_store().export_for_prompt())"

Storage Backends

The skill supports pluggable storage backends:

Graph Backend (Default)

Relationship-aware storage using CogDB - enables entity relationships, graph traversal, and richer queries.

pipx install cogdb

store = MemoryStore()  # Graph backend is default

# Add memories - entities are auto-extracted and linked
store.add_fact("User's colleague Mark is a data scientist")
store.add_fact("Mark works on the ML pipeline project")

# Graph-specific queries
related = store.find_related("Mark")  # Find all memories mentioning Mark
info = store.get_entity_info("Mark")  # Get entity graph with related entities

Storage: ~/.claude_memory/graph/

JSON Backend (Fallback)

Zero dependencies, flat-file JSON storage with thread-safe locking.

store = MemoryStore(backend="json")

Storage: ~/.claude_memory/memories.json

Project Structure

agent-memory-skill/
├── README.md              # This file
├── SKILL.md               # Detailed skill documentation
├── requirements.txt       # Dependencies (optional)
├── scripts/
│   ├── __init__.py        # Package exports
│   ├── backends.py        # Storage backends (JSON, Graph)
│   ├── memory.py          # Core MemoryStore implementation
│   └── mcp_server.py      # MCP server for Claude Code
└── references/
    ├── memory-theory.md   # Cognitive science background
    └── storage-patterns.md # Storage architecture options

Use Cases

Remember user preferences across sessions
Track project context and decisions made
Store workflows for common operations (code review, deployment, etc.)
Build personalized agents that learn from conversations
Maintain relationship context with users over time

Credits

This skill is based on concepts from the article "How Does Memory for AI Agents Work?" by Decoding AI, which provides an excellent overview of memory architectures for AI agents.

Additional references:

Liu et al. (2023). "Lost in the Middle: How Language Models Use Long Contexts"
mem0 - Open-source memory layer for AI applications

Troubleshooting

MCP Server Issues

"MCP not installed" error

pipx install mcp
# Ensure pipx bin is in PATH
export PATH="$PATH:$HOME/.local/bin"

"CogDB not installed" error

pipx install cogdb
# Or use JSON fallback by modifying mcp_server.py to use backend="json"

MCP server not appearing in Claude Code

Verify the path in config.json is absolute and correct
Check that Python can import the mcp module: python -c "import mcp"
Restart Claude Code completely (not just reload)

Python Library Issues

ImportError when importing scripts

# Ensure you're in the project directory or add to PYTHONPATH
cd /path/to/agent-memory-skill
python -c "from scripts import MemoryStore; print('OK')"

Permission errors on ~/.claude_memory

# Check directory permissions
ls -la ~/.claude_memory/
chmod 755 ~/.claude_memory/
chmod 644 ~/.claude_memory/memories.json 2>/dev/null

Storage Issues

Memories not persisting

Check write permissions to ~/.claude_memory/
For graph backend, ensure CogDB is properly installed
Try JSON backend as fallback: MemoryStore(backend="json")

Corrupted memory file

# Backup and reset
mv ~/.claude_memory ~/.claude_memory.bak
# Memories will be recreated on next use

License

MIT