Pro Workflow v3.3 Gives Claude a Persistent Memory That Compounds Over Time
Pro Workflow v3.3 adds a persistent knowledge plane to Claude Code: self-correcting memory, auto-research wikis, and multi-LLM deliberation, all on one SQLite store.
PRO- Self-correcting memory: Every correction you give Claude becomes a searchable rule, auto-loaded at session start, stored in a local SQLite database. GitHub repo
- v3.3 adds a knowledge plane: Persistent FTS5-indexed wikis that can grow themselves overnight via a budget-capped auto-research loop over web, arXiv, and GitHub sources.
- Multi-LLM council: New skill runs provider-agnostic 3-phase deliberation across Anthropic, OpenAI, OpenRouter, and Fireworks; saves transcripts as wiki pages.
- Full bundle: 34 skills, 8 agents, 22 slash commands, and 37 hook scripts across 24 Claude Code lifecycle events including PreCompact and PermissionDenied.
- Cross-agent support: Works natively in Claude Code and translates to 32+ other agents (Cursor, Codex, Gemini CLI, Windsurf) via SkillKit.
- Free and MIT-licensed: Install with two commands; optional hybrid vector retrieval and multi-LLM council require your own API keys.
Every Claude Code power user eventually hits the same wall. You correct Claude's behavior, it forgets by the next session. You explain your project conventions again. You research the same library for the third time because there is nowhere durable for the answer to land. Pro Workflow is an open-source plugin built to fix exactly that, and its v3.3 release takes the concept considerably further with a full persistent knowledge plane layered on top of its original self-correcting memory system.
The core problem it solves
You correct Claude the same way 50 times. You explain conventions every new session. Context compacts, learnings vanish, mistakes repeat. You research the same topic in three different sessions because there is nowhere durable for the answers to land. Every Claude Code user hits this wall.
Pro Workflow puts a single SQLite store underneath every session. Self-correction memory makes every correction a rule that is FTS5-searchable and auto-loaded on session start. The knowledge plane adds persistent research wikis on disk with an FTS5 shadow index, queryable from any session and optionally grown by an auto-research loop. Quality gates layer on top with LLM-powered hooks, deterministic git and secret guards, compaction-aware state, and cost tracking.
FTS5 is SQLite's full-text search extension. It lets the system index and rank corrections and wiki pages by relevance using BM25, a standard information-retrieval scoring algorithm, without needing a separate search service.
Self-correction: the compounding loop
The original mechanic is simple but powerful. When you correct Claude, the system proposes a rule, you approve it, and it gets written to SQLite. On the next session start, those rules are loaded automatically. After 50 sessions you barely correct anything. After a week of auto-research, your wiki on a topic is denser than the curated lists you started from.
Here is what that loop looks like in practice:
# Session 1
/learn-rule # You: "Don't mock the database in tests"
# Claude proposes a rule → you approve → saved to SQLite
# Session 2
# SessionStart hook auto-loads all rules
# UserPromptSubmit auto-injects top wiki hits when relevant
# Claude writes integration tests, no mock database
# Session 50
# Correction rate near zero. Wiki has 200 cited claims.This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
