HybridKV (HKV) - Bounded-memory architecture for long-context ML
HybridKV is a fixed-width state-evolution memory that replaces the linearly growing transformer key-value cache with a bounded recurrent state. Instead of storing a per-token archive, HKV maintains a single latent vector updated via a Hamiltonian-style rule. A lightweight pointer module provides token-level recall from this state with O(1) memory complexity.
Early experiments on retrieval and replay benchmarks show that HKV can reduce memory usage by approximately 20 to 30 times relative to standard KV caching while maintaining high recall fidelity. This makes HKV a candidate for long-context inference on power and memory constrained systems.
Symplectic and physics-informed foundations
CHISI began with a focus on symplectic integration and anti-dissipative dynamics in neural network optimization. The core idea is that optimizers can be viewed as discrete dynamical systems, and that structure-preserving updates can reveal hidden behavior that is invisible to standard training diagnostics.
These investigations led directly to the design of HybridKV: if an optimizer can preserve useful invariants over long horizons, a memory system can as well. HKV applies these ideas to inference-time memory by using Hamiltonian-style updates to keep a bounded state stable across long sequences.
Future directions
Ongoing work explores how bounded-state memory and structure-preserving dynamics can support tasks such as long-horizon reasoning, edge deployment, and potentially quantitative modeling where reversibility and drift control are important. If you are exploring related questions in memory architecture, symplectic integration, or physics-informed ML, feel free to reach out.