Sparse Event-Memory LM

• Last updated: 2026-08-14

SparseEventMemoryLM is a memory-efficient language-model implementation that is separate from the dense SpikingEvoTextLM / ChronoSpikeAttention path. Instead of retaining dense Q/K/V and FFN matrices as its primary memory, it uses fixed-topology INT8 CSR synapses and EvoLIF-style event state.

This page describes the current text-only implementation, but the documented extension path includes the multimodal bridge toward MemoriedSpikeNetLM and future tokenization contracts. This is not treated as a final dead-end; it is the explicit foundation for the next architecture layer.

In this section

• Concise specification — architecture, trainable state, CLI, constraints, and validation
• Detailed mathematical and memory design — rationale, mathematical model, CSR layout, dynamic tensors, sizing, execution order, and operating constraints

Implementation locations

• Model: EvoSpikeNet-Core/evospikenet/sparse_event_memory.py
• Training entry point: EvoSpikeNet-Core/examples/train_spiking_evospikenet_lm.py
• Regression tests: EvoSpikeNet-Core/tests/unit/test_sparse_event_memory_lm.py

Related pages

• Architecture details
• LLM Training System
• Documentation map
• 日本語詳細設計
• 日本語索引