Added confidence scoring to my open-source memory layer. Your AI can now say "I don't know" instead of making stuff up.

Developer remete618 has released a significant update to Widemem AI, an open-source memory layer designed for LLM agents. The core innovation is a new confidence scoring system that addresses a fundamental flaw in vector databases: they always return *something*, even when the best match is irrelevant. Widemem now classifies search results with HIGH, MODERATE, LOW, or NONE confidence levels. Developers can then choose from three operational modes: 'strict' (only returns confident facts, otherwise says 'I don't know'), 'helpful' (default, flags uncertain results), and 'creative' (admits it's guessing). This directly combats hallucination by preventing agents from fabricating answers based on low-similarity context.

Beyond confidence scoring, the update introduces several practical features for building more robust agents. The `mem.pin()` function allows developers to mark critical information—like user allergies or blood type—so it's never deprioritized or forgotten. A 'frustration detection' feature triggers a deeper search when a user indicates they've already provided information. For performance tuning, Widemem offers three retrieval modes: 'fast' (10 results), 'balanced' (25 results), and 'deep' (50 results). Crucially, the system remains fully local-first, requiring no cloud services or accounts, and is compatible with local LLM tools like Ollama and sentence-transformers under the permissive Apache 2.0 license.