Dynamic Engram: Answering from an Exact n-gram Memory Injected at Inference Time

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Instead of writing particular memories into a language model (by fine-tuning, model editing or optimized rows of an n-gram table), we train the model once to handle memories in general, and then inject any new memory on demand, with no training, no optimization and no change to the weights. Conditional memory through n-gram lookup (Engram, and the n-gram embeddings of recent open models) adds a cheap knowledge axis to transformers; its tables are learned in pre-training. Here there is no table at all. An unbounded store keeps every injected item as text and recalls the relevant ones by exact n-grams and by meaning; for each question, an exact count memory is built from the recalled items and reaches an off-the-shelf 350M-parameter multilingual model as its own token embeddings, read through Engram-style gates, and as candidates that a pointer copies at the output. The prompt holds the question alone. On memories from a life never seen in training, with the question alone in the prompt, the model answers 95.5% of associative questions correctly in the memories' language; without a memory the same data and steps give 2.7%, and with the memory emptied it answers nothing, so the facts come from the injected memory, not from the weights. The model generalizes to new memories, not yet to new forms of question (other phrasings, other languages). Against retrieval-augmented generation reading the same memories, it gives up about ten points, for a prompt two to four times shorter and an answer 1.5 to 2 times faster on two CPU threads; the gap lies mostly in questions asked in another language, while in the memories' language it is within a few points of RAG (95.5% against 100%). On Qwen3-0.6B-Base the search by meaning improves but reading transfers only in part. The model does not yet abstain when the answer is missing but similar memories are present, a gap we attribute to the training data. This is a demonstration of feasibility on small models; nothing in the method is tied to model size. This record contains the paper (PDF) and the code (library, training, evaluation and analysis scripts). The weights of the 350M model (safetensors, ONNX int8 and int4) are on Hugging Face: tetiai/dynamic-engram-350m. The training data is not distributed; it can be regenerated with the code from public datasets and synthetic memories.