PraMem

PraMem: Practice-Derived Experiential Memory for Long-Horizon Behavior Prediction

0. Environment

pip install openai tqdm jieba rank_bm25 pytz openpyxl

1. Data Preparation

Download experiment_data.json from Google Drive and place it at:

work_data/experiment_data.json

2. Build Practice Data

Step 1: Sample practice questions from each user's history.

python z_1_get_practice_source.py \
    --experiment_data ./work_data/experiment_data.json \
    --N_per_test 100
# Output: ./work_data/experiment_data.json.practice.json

Step 2: Build formatted practice inputs for each user.

python z_2_get_practice_data.py
# Output: ./work_data/practice_data/experiment_data.json.practice/{user_id}_0.jsonl

3. Build Experiential Memory

Run Self-Practice for all users. Edit z_5_practice_with_self_check.sh to set your model endpoints, then:

bash z_5_practice_with_self_check.sh
# Output: ./work_data/exp_memory/experiment_data.json.practice/{user_id}_0.json
#         ./work_data/practice_progress/experiment_data.json.practice/{user_id}_0.json

To run a single user:

python z_5_practice_with_self_check.py \
    --user_id <user_id> \
    --test_id 0 \
    --model_name http://<your-model-endpoint>/v1

4. Evaluate

Run evaluation with PraMem memory:

python evaluator.py \
    --USE_EXP_MEM 101 \
    --mem_name experiment_data.json.practice \
    --models gpt-oss-120b \
    --max-history-tokens 8000
# Output: ./work_data/results/

--USE_EXP_MEM options:

• 0 — no memory (baseline)
• 101 — PraMem