llm-user-memory

A transparent memory system for LLM that automatically maintains and uses a user profile to provide personalized AI responses.

Installation

Install this plugin in the same environment as LLM:

llm install llm-user-memory

Usage

After installation, set up transparent memory integration:

llm memory install-shell

This adds a shell function that automatically injects your user profile into every LLM interaction. Restart your terminal or run:

source ~/.bashrc  # or ~/.zshrc for zsh users

Now use LLM normally - your conversations will automatically include memory context:

llm "What should I work on today?"
# Response will be personalized based on your stored profile

llm "I just finished the memory plugin project"
# This information will be remembered for future conversations

The memory system works completely transparently. Your user profile is automatically:

• Injected as context in every prompt
• Updated in the background based on your conversations
• Stored locally in your LLM configuration directory

Features

Automatic Profile Building

The plugin automatically builds and maintains a user profile based on your conversations:

# First conversation
llm "I'm a Python developer working on machine learning projects"

# Later conversations automatically know this context
llm "What's the best way to optimize this model?"
# Response considers your Python/ML background

Transparent Operation

No need to remember special commands or flags. Once installed, the memory system works automatically:

# These all include memory context automatically:
llm "Help me debug this code"
llm -m gpt-4 "Explain quantum computing"
llm -t my-template "Process this data"

Profile Management

View and manage your stored profile:

# View current profile
llm memory show

# Clear profile and start fresh
llm memory clear

# Temporarily disable memory updates
llm memory pause

# Re-enable memory updates
llm memory resume

Background Updates

Profile updates happen in the background after each conversation, so they never slow down your interactions:

llm "I switched from JavaScript to Rust development"
# ✓ Response generated immediately
# ✓ Profile updated in background: "Updating memory..."

Privacy and Local Storage

All profile data is stored locally in your LLM configuration directory:

• No external services involved
• Profile updates use the same model you're already using
• Full control over your data

Memory Profile Structure

Your profile is stored as readable Markdown in ~/.config/llm/memory/profile.md:

# User Profile

## Personal Information
- Role: Python Developer
- Experience: 5+ years in machine learning

## Current Projects
- Working on LLM memory plugin
- Exploring transformer architectures

## Interests
- Natural language processing
- Open source development
- Performance optimization

## Preferences
- Prefers practical examples over theory
- Likes concise, actionable advice

Advanced Usage

Manual Profile Editing

You can manually edit your profile:

# Edit profile directly
$EDITOR "$(llm memory path)"

# Or use llm memory show and copy/edit content
llm memory show > temp_profile.md
# Edit temp_profile.md
llm memory load temp_profile.md

Shell Integration Details

The shell integration works by creating a function that wraps the llm command:

llm() {
    command llm -f memory:auto "$@"
}

This automatically injects the memory:auto fragment on every call.

Uninstalling Shell Integration

To remove the transparent integration:

llm memory uninstall-shell

Then restart your terminal. You can still use memory manually with:

llm -f memory:auto "your prompt here"

Development

To set up this plugin locally, first checkout the code. Then create a new virtual environment:

cd llm-user-memory
python -m venv venv
source venv/bin/activate

Now install the dependencies and test dependencies:

pip install -e .
pip install -r requirements-dev.txt

To run the tests:

pytest

How It Works

The plugin uses LLM's fragment loader system to inject profile context and monitors the conversation database to trigger background profile updates:

1. Fragment Injection: The memory:auto fragment loader reads your profile and injects it as context
2. Database Monitoring: A background process watches for new conversations in LLM's SQLite database
3. Profile Updates: After each conversation, the same model you used gets a request to update your profile
4. Transparent Operation: Shell function integration makes this completely automatic

Troubleshooting

Memory not working

Check if shell integration is active:

type llm
# Should show: llm is a function

Profile not updating

Check if background daemon is running:

llm memory status

Reset everything

llm memory clear
llm memory uninstall-shell
llm memory install-shell

Configuration

Memory behavior can be configured via environment variables:

# Disable background updates
export LLM_MEMORY_UPDATES=false

# Change update frequency (seconds)
export LLM_MEMORY_UPDATE_INTERVAL=10

# Disable memory system entirely
export LLM_MEMORY_DISABLED=true