This repository contains the official implementation for the paper: Conditional Memory via Scalable Lookup: A New Axis of Sparsity for Large Language Models.

Abstract: While Mixture-of-Experts (MoE) scales capacity via conditional computation, Transformers lack a native primitive for knowledge lookup. To address this, we explore conditional memory as a complementary sparsity axis, instantiated via Engram, a module that modernizes classic $N$-gram embeddings for $\mathcal{O}(1)$ lookup.

Key Contributions:

• Sparsity Allocation: We formulate the trade-off between neural computation (MoE) and static memory (Engram), identifying a U-shaped scaling law that guides optimal capacity allocation.
• Empirical Verification: Under strict iso-parameter and iso-FLOPs constraints, the Engram-27B model demonstrates consistent improvements over MoE baselines across knowledge, reasoning, code and math domains.
• Mechanistic Analysis: Our analysis suggests that Engram relieves early layers from static pattern reconstruction, potentially preserving effective depth for complex reasoning.
• System Efficiency: The module employs deterministic addressing, enabling the offloading of massive embedding tables to host memory with minimal inference overhead.

The Engram module augments the backbone by retrieving static $N$-gram memory and fusing it with dynamic hidden states. The architecture is shown below (drawio provided):

We recommend using Python 3.8+ and PyTorch.

pip install torch numpy transformers sympy

We provide a standalone implementation to demonstrate the core logic of the Engram module:

python engram_demo_v1.py

⚠️ Note: The provided code is a demonstration version intended to illustrate the data flow. It mocks standard components (like Attention/MoE/mHC) to focus on the Engram module.

The use of Engram models is subject to the Model License.

If you have any questions, please raise an issue or contact us at service@deepseek.com.