V1 Progress — May 15, 2026

What shipped

Site data sync (May 13–15)

The full data pipeline from WordPress to dashboard is live:

• WordPress plugin (memberintel-connect): 6 REST endpoints exposing members, transactions, subscriptions, memberships, stats, and system info from MemberPress sites. Fixed memberships query (WP CPT, not custom table), subscriptions (no expires_at column), mp_version (get_file_data instead of get_option), and member_id: 0 filtering.
• Backend sync service: Paginated fetch from all WP endpoints, upsert into PostgreSQL with UniqueConstraint(site_id, mp_id) dedup. Five new DB models: Member, Transaction, Subscription, Membership, SiteStats.
• SPA dashboard: Stat cards (members, MRR, revenue, churn), transactions table, memberships table, sync button.
• SPA settings: Connected sites list with sync/disconnect, account info, plan section.
• Multi-site support: Topbar dropdown for switching between sites, activeSiteId in persisted app state.
• Marketing site auth awareness: SPA sets .membersintel.com cookie, marketing /connect/ page redirects logged-in users to the app.
• Cloud Run: Backend deployed at memberintel-api-staging.
• Cloudflare Pages: SPA at app.membersintel.com, marketing at membersintel.com.

Chat improvements (May 15)

Three critical fixes to the chat pipeline:

1. Site data context injection: When a user has an active site, their stats, memberships, and recent transactions are injected into the system prompt. The AI can now answer site-specific questions like “How many members do I have?” and “What’s my MRR?”
2. Conversation history: The LLM now receives the last 10 messages from the conversation, enabling multi-turn dialogue. Previously each message was standalone with no memory of previous turns.
3. Voyage query embedding fix: embed_query() now uses input_type="query" instead of input_type="document", improving retrieval quality per Voyage’s documentation.

Customer brain spec (May 15)

Design spec for the per-customer persistent context system: 2026-05-15-customer-brain-design.

Four document types that make the AI advisor remember and improve:

Document	Owner	Purpose
SOUL	User	Preferences, thinking style, common questions
BIBLE	Site	Foundational truth about the site — what it is, who it serves
HEARTBEAT	Site	Current state, auto-generated from sync data
MEMORY	User+Site	Running notes from conversations, retrieved by vector search

SOUL + BIBLE + HEARTBEAT are always-on context in every system prompt. MEMORY is retrieved by vector search when relevant.

---

What’s next

Phase A: Schema + HEARTBEAT (replaces current site-context injection)

• Alembic migration: user_souls and site_contexts tables
• generate_heartbeat() template function (replaces raw SQL build_site_context())
• generate_initial_bible() template from site URL + memberships
• Hook HEARTBEAT generation into sync_site()
• Update system prompt to inject SOUL + BIBLE + HEARTBEAT

Phase B: LLM tool calls (update_customer_brain)

• Add Anthropic tool use support to call_stream() and call()
• Implement update_customer_brain tool handler
• SOUL/BIBLE updates (read-modify-write), MEMORY creation (new brain_entry)
• Tests for tool call interception and persistence

Phase C: Memory retrieval + API

• Wire collection='memory' search into chat pipeline
• Add brain API endpoints (GET/PUT soul, GET/PUT bible, GET heartbeat)
• RAG re-ingestion from Hive Mind (ensure current data)