Pluggable, removable fact memory for a frozen model — running entirely in this tab. Engram fact cartridges are a few MB each and mount in milliseconds: two hold every Premier League result of a season (2025-26 and 2024-25, both after the base model's training cutoff, stacked on a LoRA behaviour cartridge that teaches answer format); one holds the population of the world's 16,000 largest cities. Ask, swap, unmount — the knowledge is provably gone. The Chat workspace puts an instruct model in front of the same memories: ask in plain language, watch it write the memory's key, and unmount to make it forget.
Fetches the cartridges (~300 MB) and the frozen base model (~1.2 GB, all cached in this browser after the first visit) from Hugging Face.
A fact cartridge is built against the exact stack it runs on: the season cartridges target the base merged with the format LoRA; the cities cartridge targets the bare base; the chat workspace uses an instruct base (Qwen3-0.6B chat, an extra ~1.5 GB fetched on first use) with all three memories rebuilt against it. Switching workspaces reloads the model from memory (~10 s) with the right stack.
Ask in plain language. The model decides what to look up and writes the memory's exact key; the mounted cartridge answers it; the model turns the value into a sentence. Unmount the memory and it forgets — and says so.
Cartridges & base model served from huggingface.co/lewisdog/engram-md-demo. Results data: football-data.co.uk. City data: GeoNames (CC-BY 4.0). Base model: Qwen3-0.6B-Base (Apache-2.0). Runtime: wllama (engram fork). Nothing you type leaves this tab. Contact: lewis@engram.md.