Self-correcting memory for Claude Code and 32+ agents

What it does

Pro Workflow is a skill bundle and plugin that drops a SQLite database under your AI coding sessions. It captures corrections as FTS5-searchable rules, auto-loads them at session start, and builds persistent research wikis that can grow via an optional, budget-capped auto-research loop. It ships with 34 skills, 22 slash commands, and hook scripts across 24 events to gate commits, track costs, and audit MCP overhead.

The interesting bit

The project treats your session history as a compounding asset rather than disposable context. Corrections harden into rules, wikis auto-populate through BFS research with a kill switch, and a multi-provider LLM council can deliberate on decisions with transcripts saved back to the wiki. It is essentially an opinionated local knowledge layer pretending to be a plugin.

Key highlights

• Self-correcting memory: every approved correction becomes a rule indexed in SQLite and injected on SessionStart.
• Persistent knowledge plane: FTS5-indexed wikis with optional hybrid retrieval (BM25 + vector RRF) and an auto-research loop capped by budget.
• Cross-agent support: native Claude Code plugin plus SkillKit translation for Cursor, Codex, Gemini CLI, and 26 others.
• Quality gates and observability: LLM-powered hooks, cost tracking, compaction-aware state, and MCP overhead auditing.
• Multi-provider deliberation: llm-council runs a 3-phase debate across Anthropic, OpenAI, OpenRouter, or Fireworks and persists the transcript.

Caveats

• SkillKit installation currently requires --force because its security scanner flags standard Node patterns; the fix is tracked at skillkit#129.
• Hybrid retrieval and embedding need an external API key (OPENAI_API_KEY or VOYAGE_API_KEY), so the fancier search is not fully local.
• The comparison table notes that some competing bundles ship more raw skills, though Pro Workflow claims tighter integration around memory and hooks.

Verdict

Worth a look if you spend your first ten minutes of every Claude Code session re-establishing ground rules. Skip it if you prefer ephemeral, stateless chats or do not trust a local SQLite file to outlast your dotfiles.