Ex8-ca

Self-hosted vector memory (Qdrant + TurboQuant compression) with auto-discovered knowledge realms, semantic search across conversations and ingested documents, and a slumber pipeline for memory consolidation. v1.1.0 adds optional per-session injection dedup with char-budget gate, decision/evaluation memory-type promotion (slower decay for load-bearing turns), prompt-injection sanitization on read, configurable system-prompt template, export/import memories as JSON, and a /memex8 slash command. New memory behaviours are opt-in; /memex8 export|import is confined to $HERMES_HOME/plugin-data/memex8 unless enable_export_import=true. Disclosure — auto-stores a truncated copy (first 300 chars) of every user and assistant turn, plus per-message previews of the whole transcript at session end and before context compression, to the memex8 server at MEMEX8_BASE_URL (self-hosted; default http://localhost:8080) using MEMEX8_API_KEY.