EdgeRock open-sources constant-memory reasoning: O(1) history replaces growing KV cache
eyishazyer · x · 2026-11-04
Rockmatrix AI has released EdgeRock, a multimodal model line that does "strong reasoning at constant memory": O(N) generation with O(1) persistent history, eliminating a growing KV cache.
- The key mechanism is an admission step: every 256 tokens the model decides which new details earned an anchor, keeping the context file useful even at session 100.
- The author stresses this capability is learned in training, not bolted on via external files and rules like typical agent memory setups.
- Weights, inference/eval code, paper and whitepaper are all out; GitHub repo is Apache-2.0 licensed and at 993 stars, with models on Hugging Face.
Related event: EdgeRock-4B Open-Sources Constant-Memory Inference for LLMs(3 posts)→
More from Models
- Claude Code fixes game copyright refusal rule in 2.1.317: naming a game now works — apples_jimmy · 2026-11-04
- Scenario-based picks: how to choose among 7 small language models — goyalshaliniuk · 2026-11-04
- LFM2.5-1.2B and SmolLM3-3B: on-device and research picks among 2026–27 SLMs — goyalshaliniuk · 2026-11-04
- Phi-4-mini and SmolLM3-3B specs detailed in 2026–27 SLM roundup — goyalshaliniuk · 2026-11-04
- Microsoft Phi-4-mini: a 3.8B model with 128K context and 52.8 MMLU-Pro — goyalshaliniuk · 2026-11-04
- IBM Granite 4.2 3B lands on list of 7 small language models worth exploring — goyalshaliniuk · 2026-11-04