Governed white-box memory for LLM agents: every fact carries a mechanically-verified evidence span, every rejected hypothesis is documented. Local-first (SQLite+FTS5), 1000+ tests. LV-Eval-256k 91.94% / NoLiMa-32k literal 96.43% on a local 7B — negative results disclosed.
agent memory sqlite mcp memory-management obsidian hallucination rag ai-agent llm long-context langchain local-llm ollama context-window qwen longmemeval vector-database-alternative nolima lv-eval
-
Updated
Jul 11, 2026 - Python