Why it’s interesting

Human-like memory: consolidates context across sessions; periodic cleanup keeps long-term state sane.

Self-learning (Q-utility): prioritizes what to remember/retrieve based on usefulness.

Self-reflection: validates and corrects its own reasoning traces.

Languages: English + Russian today; CN/ES planned.

Real-world test Ran in an air-gapped setup processing 15–20 MB PDFs (policies/protocols) and answering multi-hour dialogue queries without any external APIs.

Tech notes

FAISS for ms-level semantic search

SQLite for offline storage

Async pipeline; streaming responses

Rule/logic parser for “if A and B then C” style queries

Web UI + REST API (/api/chat, /api/upload); modular LLM backend (currently Ollama)

Repo (public marketing repo; core code abstracted for privacy): https://github.com/vac-architect/vac-agi