---

The components of MegaContext work together to enable Context Focus.


Details

  • Architecture — dig into the details of the system architecture.
  • GistNet — 32→1 hierarchical gists aligned with the base embedding space.
  • LensNet — dual cross-attention scorer providing signed utilities per Working Context entry.
  • Focus Allocator — greedy, hysteresis-aware application of LensNet utilities.
  • Positional Encoding — global index and LOD-aware positional strategy for frozen and co-trained models.
  • Multimodal MegaContext — image gist hierarchies, multimodal positional encoding, and mixed-LOD integration.
  • Multi-headed Focus — advanced focus strategies (multi-head and staging contexts) extending the baseline loop.
  • Runtime Loop — streaming ingest, focus adjustment, and base LLM decode.
  • Training & Operations — counterfactual labeling, alternating optimization, and telemetry.
  • Performance Sketch — expected compute/storage envelopes at various scales.