J-Space-Cognition-Suite-V3.6

What it solves

The J-Space Cognition Suite addresses insufficient working-memory management during deep reasoning, long-horizon tasks, tool use, verification, and recovery. It organizes an agent’s accessible working representations through selective loading, a broadcast hub, dense tracks, bridge-before-conclusion reasoning, metacognitive control, and other mechanisms — all applied at inference time without modifying model weights.

Who it’s for

• Developers building agents on DeepSeek Harness or any native Skill loader.
• Users handling multi-step, multi-file, multi-turn, or long-persistent-task workloads.
• Teams seeking to boost model reasoning reliability without retraining.

Key features

• Model-agnostic: Works purely at inference time; weights and training remain unchanged.
• Selective loading: Dynamically loads the lightest sufficient set of modules.
• Optional controller: jspace.py externalizes long-task state to .jspace/, keeping solution choice with the model.
• Comprehensive docs: Quick start, operating modes, mechanism explanations, benchmarks, and a Chinese README.