mirror of
https://github.com/wangcanran/MyMerchMind.git
synced 2026-11-22 16:37:33 +08:00
Core Features
- New ExperienceStore module (ecommerce_agent/memory/experience_store.py)
- Structured storage: title/narrative + search_keys + evidence
- Search keys derived from system-collectable fields only (season/category/price_tier/channel/return_signal/inventory_signal/conversion_signal/tags/trend_keywords)
- Key-overlap scoring for retrieval, no embedding dependency
- State machine: draft -> active -> archived
- Auto-draft generator (ecommerce_agent/memory/experience_draft_generator.py)
- High-return SKUs -> draft
- Slow-moving/overstocked SKUs -> draft
- New product planning approve_new -> draft
- Large pricing deviations -> draft
- System captures numeric facts, narrative left for human refinement
- Orchestrator integration (orchestrator.py)
- Retrieve relevant experiences at run() start, inject into memory_snapshot
- Auto-generate drafts at run() end, persist to store
- Support experience_store_path via env var or parameter
- REST API (api/main.py)
- GET /api/experiences - list experiences (filter by status)
- POST /api/experiences/search - search by system fields
- POST /api/experiences - manual entry (active directly)
- POST /api/experiences/{id}/approve - promote draft to active
- POST /api/experiences/{id}/archive - archive experience
- GET /api/experiences/{id} - get single experience
Design Principles
- Experience = reference, not hard constraint; default advisory
- Search only uses system-collectable fields; narrative excluded from hard filtering
- Auto-drafts require human or rule promotion before becoming active
- Coexists with existing FeedbackMemoryStore
Next Steps
- Add embedding semantic layer for rerank (currently pure key-overlap)
- Bidirectional sync with LightRAG (ingest high-confidence experiences into graph)
Co-authored-by: Cursor <cursoragent@cursor.com>