Changelog

All notable changes to ChorumAI will be documented in this file.

[Unreleased]

Deterministic Grounding: Implemented a safety layer that verifies new learnings against recent conversation history to prevent “poison-pill” injections.
- Added verifyReference check before committing memories.
- Introduced Source Tagging (Provenance) to track which conversation and messages contributed to a memory.
Confidence Rescoring: Introduced a dynamic confidence scoring formula for projects.
- Formula: (Interaction*0.3 + Verification*0.4 + Consistency*0.2 + Decay*0.1) * 100

Added scripts/test-grounding.ts and scripts/test-confidence.ts for validating memory logic.

Co-occurrence Scoring: Items frequently retrieved together with positive feedback now receive a relevance bonus (up to +0.10), leveraging existing co-occurrence tracking data.
Intent-Adaptive Score Thresholds: Relevance thresholds are now per-intent — debugging casts a wider net (0.25), generation demands precision (0.40), greetings set a high bar (0.50).
Confidence-Gated Injection: selectMemory() applies intent-specific minimum score thresholds before injection.
Dynamic Weight Shifting: Scoring weights shift based on conversation context — deep conversations boost recency, code-heavy queries boost domain matching, history references boost semantic similarity. All weights re-normalize after shifting.

Expanded Debugging Detection: Added 10 new keywords (trace, stack, exception, undefined, typeerror, null pointer, logs, breakpoint, segfault, panic) for more accurate debugging intent classification.

High-Usage Promotion: Learning items with ≥10 retrievals are automatically promoted, guaranteeing inclusion in Tier 1/2 compiled caches regardless of decay.
New promotedAt column on project_learning_paths.
Promotion runs automatically before cache recompilation.