Show HN: Replacing Turn-Based LLM Agents with SQLite and System One Models
Category: ai-ml
Tags: coding-agents, sqlite, retrieval, tree-sitter, system-one, mcp, code-indexing
Score: 6.5/10 (Innovation: 8, Technical: 7, Documentation: 5, Utility: 6)
unified-brain is a persistent SQLite-based memory and perception layer for coding agents that replaces turn-based autoregressive exploration with hundreds of bounded yes/no judgments from a 'System One' judge model (Jev). It indexes full git history symbol-by-symbol via tree-sitter, then retrieves, filters, and synthesizes answers with a single writer LLM call, claiming frontier-agent accuracy at a fraction of cost and tokens. The core idea—decoupling non-autoregressive evidence gathering from synthesis—is genuinely interesting, but it depends on a proprietary judge model and the energy/accuracy claims are largely unmeasured.
Target audience: backend devs and AI engineers building or using coding agents
Repository: https://github.com/DevLoops-ai/unified-brain · Python · AGPL-3.0 · 29 stars
View on Hacker News