lewisdog/engram-md-demo

Engram cartridges — pluggable fact memory for a frozen model

Engram cartridges are small, detachable n-gram fact memories (DeepSeek-style hash-keyed lookup tables) that mount onto a frozen base model in milliseconds, add exact-recall knowledge the base model cannot have, and unmount leaving the base byte-identical. Built with aoa-engram and runnable in llama.cpp (engram branch) natively or in the browser via a wllama fork.

Files

File	What it holds	Recall
qwen3-0.6b-base-f16.gguf	The exact frozen base (F16 conversion of Qwen/Qwen3-0.6B-Base) all cartridges are fingerprint-bound to	—
cartridge-premier-league-2526.gguf	Every 2025-26 Premier League result (380 matches)	97.9%
cartridge-premier-league-2425.gguf	Every 2024-25 Premier League result (380 matches)	97.9%
format-lora-2324.gguf	LoRA behaviour adapter: answer format + stopping, trained on the 2023-24 season	—
cartridge-cities-16k.gguf	Populations of the world's 16,000 largest cities (GeoNames)	100% (200-sample)
engram.md	Machine-readable manifest: key formats, fact counts, base-model fingerprints	—
demo-facts.json	Sample prompts with ground-truth answers for the browser demo	—

Both Premier League seasons finished after the base model's training cutoff — the frozen model provably cannot know these results. The two football cartridges are trained against the base merged with the format LoRA (a fact cartridge must be built against the stack it runs on); the cities cartridge is trained against the bare base.

Query format is exact (see engram.md), e.g.:

Newcastle v Liverpool | result |   ->  2-3 (Liverpool)
Nanjing, CN | population |         ->  9314685

Data sources & licenses

• Base model: Qwen/Qwen3-0.6B-Base, Apache-2.0.
• Football results: football-data.co.uk (free historical results data).
• City populations: GeoNames cities15000, CC-BY 4.0.