ENGRAFT
Early-stage researchENGRAFT is an experimental technique, a few weeks old. What you read here are first measurements on one model, not a product.

About

An open method, built in the open

ENGRAFT is an independent research project on writing facts into the n-gram memory of large language models. The method, the code that measures it and every result file are public.

Who

ENGRAFT is developed by fulvian, working with Claude Code. The operator set the goals, approved every design step, ran the hardware and read every result. Design, implementation, adversarial review and independent verification were done by separate model instances. We say it on every page that matters because you should know it before reading the numbers.

The experiments run on one machine: an AMD Ryzen AI MAX+ 395 with 128 GB of unified memory, running Qwen3.8-Flash-Next through a fork of llama.cpp.

The names

ENGRAFT stands for ENgram GRadient Routing-Aware Fact Transplant. It is not related to ENGRAFT (CCS 2022, Byzantine consensus) or to engraft.dev; the domain of this site, engraft-engram.dev, pairs the method with the table it writes into.

Licenses

  • Code: Apache License 2.0.
  • Overlays produced against a Qwen3.8-Flash-Next GGUF are derivative works of that model and fall under the Qwen Community License 1.0.
  • The Quail corpus: CC BY 4.0 on our own contributions only. It is inspired from memory by Philip K. Dick's short story We Can Remember It for You Wholesale (1966); no text of the story was copied and no right in that work is granted (PROVENANCE.md).
  • The engine fork: MIT, like llama.cpp.

Citing

The technical report (version 3, PDF) has a BibTeX entry on its page; the repository can be cited through its CITATION.cff. An arXiv version will follow; its identifier will be added here on the day it appears.

  • Engram (DeepSeek): Conditional Memory via Scalable Lookup, Cheng et al., arXiv:2601.07372. The table design this method edits.
  • User as Engram, Bojie Li, arXiv:2606.19172. Per-user memory as local edits of a hash-keyed table on small Engram models; closest in spirit.
  • Engram Adapter, Hou et al., arXiv:2608.29327.
  • Memory Grafting, Cheng et al., arXiv:2605.20948.
  • ngram-knowledge-injector, ortegaalfredo/ngram-knowledge-injector: patches the same table with overlay files, computing the rows differently.
  • llama.cpp PR 27742 added the architecture the engine fork builds on.

Contact

Questions, corrections and proposals go to the issues of the repository. A correction that changes a number is credited on the page that carries it.