Dual-Brain Memory Guardian

Problem: Rules and Experience Drift Apart

AI agents often repeat the same mistakes across long conversations or recurring tasks. A user correction may not influence the next attempt, completed work may lack structured reflection, and past pitfalls are usually retrieved by brittle keywords instead of fuzzy semantic matching. At the same time, high-confidence behavior constraints must not be overwritten by noisy recall.

How It Works: Rule Brain + Experience Brain

Dual-Brain Memory Guardian uses a dual-brain contract:
- Left brain (Markdown): stores explicit rules, durable preferences, and behavior constraints in files such as SKILL.md, operations.md, learning.md, and boundaries.md.
- Right brain (Pinecone): stores corrections, reflections, and task experience, with semantic recall using PINECONE_API_KEY and multilingual-e5-large.

Key trigger points include memory:session-start, memory:auto-session-start, memory:on-correction, memory:on-task-complete, and memory:mark-promoted. The session starts by loading relevant memory, corrections are captured as events, non-trivial tasks end with a reflection, and only then may experience be promoted into rules according to learning.md.

Boundaries and Notes

The skill keeps rule memory under ~/dual-brain-memory-guardian/ and stores experience memory in Pinecone. The conflict order is fixed: project Markdown > domain Markdown > global Markdown > Pinecone recall; vector recall should not outrank explicit contracts. It requires Node.js 20+, npm, and the Pinecone SDK, and it should not store sensitive secrets or perform destructive rewrites of uncertain content.