pro-workflow: a persistent memory and self-correction layer for AI coding agents with auto-growing research wikis

What it solves

Pro Workflow addresses the "memory loss" and repetitive correction cycle common in AI coding assistants like Claude Code and Cursor. It prevents users from having to explain the same conventions or correct the same mistakes across different sessions by providing a durable, searchable memory system.

How it works

The project implements a single SQLite store that acts as a persistent layer underneath AI sessions. It uses several key mechanisms:

• Self-Correction Memory: Captures user corrections as rules that are automatically loaded at the start of every new session.
• Knowledge Plane: Creates FTS5-indexed research wikis on disk that can be queried via BM25 retrieval or hybrid vector search. It includes an auto-research loop that can grow these wikis autonomously using budget-capped BFS (Breadth-First Search) over web, arXiv, and GitHub sources.
• Quality Gates: Employs LLM-powered hooks and deterministic guards to scan for secrets, validate commits, and manage context compaction.
• Cross-Agent Translation: Uses SkillKit to translate its suite of skills and commands into the native formats of over 32 different AI agents (including Cursor, Gemini CLI, and OpenHands).

Who it’s for

Developers using AI coding agents (specifically Claude Code, Cursor, and others) who want to build a long-term knowledge base for their projects and eliminate repetitive prompting.

Highlights

• 34 Skills & 22 Commands: A comprehensive toolkit for everything from /develop (multi-phase feature building) to /smart-commit (quality-gated commits).
• Auto-Research Loop: An opt-in, budget-capped system that autonomously expands research wikis.
• Multi-LLM Council: A provider-agnostic deliberation system that uses multiple LLMs to reach a consensus on high-stakes decisions.
• Hybrid Retrieval: Combines BM25 and vector RRF for efficient knowledge retrieval from the internal wiki.
• Extensive Hook System: 37 scripts across 24 events (e.g., SessionStart, UserPromptSubmit) to automate agent behavior.