Memory and context management improvements

[bug-ops]

Overview

Competitive analysis identified gaps in memory and context management subsystems. This epic tracks improvements across three priority tiers.

P0 — Critical

• Implement embedded vector search fallback via SQLite BLOB storage #741 — Embedded vector search fallback (SQLite BLOB-based, zero external dependencies)
• Improve token estimation accuracy for multibyte text #742 — Improve token estimation accuracy (chars/4 + safety margin)
• Add credential scrubbing to LLM context pipeline #743 — Credential scrubbing in LLM context pipeline

P1 — Important

• Add MMR re-ranking to hybrid recall #744 — MMR re-ranking for recall diversity
• Add temporal decay scoring to semantic recall #745 — Temporal decay scoring for recency bias
• Implement adaptive chunked compaction #746 — Adaptive chunked compaction
• Add compact skills prompt mode for small-context models #747 — Compact skills prompt mode for small-context models

P2 — Nice-to-have

• Autosave assistant responses to semantic memory #748 — Autosave assistant responses to semantic memory
• Implement memory snapshot export and import #749 — Memory snapshot export/import
• Add LLM response cache with SQLite backend #750 — LLM response cache

Reference

• Analysis: .local/plan/memory-context-improvements.md