Engram Conditional Memory (N-gram Hash Lookup + Offload/Prefetch)

What problem does it solve?

Engram Conditional Memory provides fast, deterministic memory lookups for language models by hashing N-gram contexts to embedding vectors, with optional CPU offload and asynchronous prefetch to scale memory usage.

Core Features & Use Cases

• Deterministic multi-head N-gram hashing for memory retrieval
• Tokenizer compression integration to reduce vocabulary size efficiently
• Phase 1 encoder-competition and Phase 2 layer-augmentation to fuse memory with backbone processing
• CPU offload plus async prefetch to scale large embedding tables without exhausting GPU memory
• Flexible integration with gating, depthwise causal convolution, and residual fusion for stable training

Quick Start

Run Engram memory with default configuration and enable CPU offload to observe prefetch-driven memory retrieval.