Aurra ships automatic memory supersession for AI agents with LLM confidence gating and per-category safety exclusions

Aurra released bi-temporal versioning with automatic supersession detection (Level 2), complementing yesterday's manual API. When a new memory is written with auto_supersede=true, an LLM classifier compares it against the top 3 semantically similar existing memories and returns one of three verdicts: 'supersedes', 'refines', or 'independent', each with a confidence score (0.0–1.0). The system acts only if verdict == 'supersedes' AND confidence >= 0.85.

The classifier uses a 4,771-character system prompt with calibration scaffolding, linguistic signal hints ('switched', 'moved', 'cancelled' for supersession; 'also', 'additionally', 'second' for independence), and tiebreaker rules that default to NOT superseding when uncertain. Memories in customer-configured excluded categories—defaulting to ['health_medical', 'legal_status']—never enter the LLM; the category gate runs in code before any classifier call.

Testing on 121 hand-labeled cases showed claude-haiku-4-5 achieved 100% precision on gated supersedes verdicts (55/55 correct), 91.7% recall, 93.4% overall accuracy, at ~$0.0005 cost per classification. The 8 cases haiku missed were false negatives (recoverable via manual supersede API); one false positive in sonnet testing would have been blocked by the health_medical category gate in production.

API surface and SDKs ship today (aurra==0.3.1 Python, aurra@0.2.1 npm). Live classification is gated by a server-side environment variable; calls currently return skipped_reason: 'level_2_disabled_by_env' while final production validation runs, with the flag flipping on later this week.