v0.9.20 — ASMR Agentic Search Mode

What's New

Agentic Search Mode (ASMR-inspired)

Three parallel LLM search agents replace embeddings for memory retrieval. Achieved 98.4% Recall@5 on full LongMemEval-S (500 questions).

• Direct Seeker — exact match, literal facts, names, numbers
• Inference Engine — implicit connections, context, implications
• Temporal Navigator — timelines, chronological ordering, state changes

Smart Context Loading (3-bucket)

• Bucket 1: Project memories (DB-filtered by source_ref)
• Bucket 2: Profile memories (preferences, guidelines, gate-rules)
• Bucket 3: Recent memories (14 days, same project only)
• Relevance-sorted by feedback signals (relevant_count, noise_count)
• 700K char budget cap

Categorized Observer Extraction

Facts now tagged with [PERSONAL], [PREFERENCE], [EVENT], [TEMPORAL], [UPDATE], [DECISION] for better agent reasoning.

Interactive Setup

• cems setup now has checkbox-style IDE selection
• Search mode selection (auto vs agentic)
• cems update preserves custom credentials (CEMS_SEARCH_MODE, CEMS_TEAM_ID)

Model & Performance

• Switched to Gemini 2.0 Flash (3x cheaper, faster, same 1M context)
• Hook results reduced to 3 in agentic mode
• 8s server-side timeout prevents orphaned LLM requests

Bug Fixes

• Fixed _doc_row_to_dict missing relevant_count/noise_count (relevance sorting was inert)
• Fixed case-sensitivity mismatch in project filtering
• Fixed asyncio.get_event_loop() deprecation
• Fixed cems update wiping custom credential settings