Repository for the paper "mem-agent: Equipping LLM Agents with Memory Using RL".

1. Copy environment file:

   make copy-env

   and add the relevant values w.r.t. what you want to do in the repo. The OPENAI_API_KEY, WANDB_API_KEY and WANDB_PROJECT are required for training and the OPENROUTER_API_KEY is required for evaluation.

2. Check if uv is installed and install if needed:

   make check-uv

3. Install dependencies (root and agent project):

   make install

1. Prepare memory from instances:

   make setup-memory

2. IMPORTANT: For the training of Qwen2.5 and upwards, use this make target to remove a potential error vLLM might throw mid-training:

   make remove-vllm-error

3. Review configuration:

   • Edit config.json to adjust model and hyperparameters.

4. Start training:

   make train

Notes:

• On macOS, GPU-specific packages are automatically skipped during install and train.

Runnign the evaluation is very simple, uses OpenRouter by default and thus accepts OpenRouter model names.

# Basic (set your preferred model)
make eval MODEL=qwen/qwen3-8b

# Use vLLM or LMStudio for local inference
make eval MODEL=qwen/qwen3-8b USE_VLLM=1

Evaluation reads data from data/eval/.