Memory models demo

Two small open models for an AI assistant's long-term memory. kev-mem-0.8b decides whether a new memory replaces an old one. qwen3.5-4b-engram answers questions from retrieved memories. Everything below is a real model output, produced on a Kaggle T4. To try your own inputs, see Run it yourself.

By Abhash Chakraborty · Models: kev-mem-0.8b · qwen3.5-4b-engram
0.954mean ROC AUC on stated updates, held-out tests
0.834Jev on the same tests · Kev-0.8B as released 0.852

Pick an example
New memory
Existing memories next to it

p_update = (P(relation = updates) + P(outdated)) / 2, one forward pass for all questions. The engine's default gate is 0.70: in testing it never replaced a memory that was still true, but it misses some real changes. Move the slider to see the trade-off.

0.766held-out LoCoMo judge accuracy · best other open 3-4B model 0.662
8 / 8correct on the examples below (short answers, dates resolved, current value preferred)

Free GPU demo with your own inputs (Kaggle)

Open memory-models-demo, click Copy & Edit, turn on a GPU T4 and Internet, then Run All. The last cell prints a public gradio.live link with both models (about 2 s per decision, 1-2 s per answer).

Run locally

git clone https://huggingface.co/abhash-chakraborty/kev-mem-0.8b
cd kev-mem-0.8b/demo && pip install -r requirements.txt gradio && python app.py

git clone https://huggingface.co/abhash-chakraborty/qwen3.5-4b-engram
cd qwen3.5-4b-engram/demo && pip install -r requirements.txt gradio && python app.py

kev-mem runs on a CPU in a few seconds per decision (about 4 GB of GPU memory, if used). The 4B model needs about 10 GB of GPU memory (8.6 GB in 4-bit), or 30-60 s per answer on a CPU.

Benchmarks

kev-mem System-1 benchmark · Kaggle Benchmark: memory update detection · engram vs public models · Kaggle Benchmark: memory QA