## Transparent in-memory compression for Linux (the "macOS gap")
→ Sourced into docs/action-plan.md (Milestone 1) and docs/roadmap.md (Stage 1).
→ Kernel module scaffold exists. Remaining: PTE marking, page fault handler, idle tracking, policy, parallel decompression.

## VRAM compression for AI workloads
→ Sourced into docs/action-plan.md (Milestones 3-4) and docs/roadmap.md (Stage 2).
→ Research complete: mm/ has zero VRAM visibility (research/014), TTM is insertion point, MoE has 70-90% cold weights (research/017), tiered VRAM strategy defined, no existing offloading system uses compression.

## Compression advisor / auto-selector
→ Complete. Heuristic decision tree in src/lib/advisor.c. Picks best algorithm per page type.

## CXL memory with inline compression
→ Sourced into docs/roadmap.md (Stage 3). CXL Type-3 works today with Stage 1 module — no new infrastructure needed.

## Cross-process page deduplication + compression
→ Still in design. Post-Stage 1. KSM merges identical pages; MiniMem could compress similar pages using delta encoding. Research needed on efficient similarity detection.

## Standalone library (libminimem) for AI ecosystem
→ Sourced into docs/action-plan.md (Milestone 2).
→ Package algorithm library as standalone .so/.a with stable API. Target users: llama.cpp, vLLM, Ollama, self-hosting tools.
→ Key value: FP16 compressed 2:1 with zero quality loss — enables running larger models on same hardware.

## DSC-inspired lossless predictor ("DSC-Lite")
→ Sourced into docs/roadmap.md (Stage 4) and research/015.
→ MMAP prediction + ICH dictionary for 32/64-bit words, no quantization. Estimated 1.5-2.5:1 on structured pages. Needs implementation.

## Parallel cluster decompression
→ Sourced into docs/action-plan.md (Step 4) and research/016.
→ 4.5-7× latency reduction on 32-page clusters. Swap readahead provides clustering. mmu_gather for batched TLB flush.

## Sparse activation map for MoE inference
→ Sourced into research/017 and docs/action-plan.md (Milestone 3).
→ MoE router is a natural activation map. Tiered VRAM: hot (uncompressed) → warm (compressed VRAM) → cold (compressed RAM) → frozen (compressed NVMe).