…ory → FullMemory)

Introduces CoreMemory (zero-config, no LLM), SmartMemory (echo/categories/graph),
and FullMemory (scenes/profiles/tasks) with CoreSQLiteManager, content-hash dedup,
query cache, presets system, benchmark command, and 82 new/updated tests.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

…elf-improvement loop

Adds skills as first-class citizens to Engram's memory architecture:
- Skill schema (SKILL.md format), store, executor, outcome tracking, and discovery
- Trajectory recording and persistence for tracking agent task episodes
- Skill miner that compiles successful trajectories into reusable skills
- zvec vector store backend with graceful fallback chain
- Centralized SHA-256 hashing (content, trajectory, skill signature)
- 6 new MCP tools for skill and trajectory management
- SmartMemory gains skill search/apply; FullMemory gains trajectory recording and mining
- 71 new tests, all 105 tests passing

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

…DIA provider updates

- Add unified enrichment module with multi-modal encoding pipeline
- Enhance skill system with structure module, improved executor/miner/outcomes
- Update memory architecture (smart memory, main memory) for enrichment integration
- Add NVIDIA embeddings and LLM provider improvements
- Update LongMemEval benchmark runner
- Add structural and unified enrichment test suites

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

…ment optimizations

Add cross-encoder reranking stage using NVIDIA NIM API for improved
retrieval accuracy. Add nemotron-embed model support with proper
modality list handling in batch embeddings.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>