Grocery Memory

This ensures the planner always works with up-to-date preferences, staples, and fridge state.

Memory Files

All files live in /workspace/group/grocery/. Create the directory if it doesn't exist.

How to Analyze Deliveries

1. Fetch order history

# Use rohlik MCP tools — do NOT scrape the website

Call mcp__rohlik__get_order_history to get recent orders. Then call mcp__rohlik__get_order_detail for each order to get item-level data. Process orders one at a time to avoid context overflow.

2. Extract preferences

For each order, record:

• Items and quantities — what was bought and how much
• Brands — track which brand is chosen when alternatives exist
• Categories — group items (dairy, meat, vegetables, bakery, pantry, frozen, drinks, household)

After processing all orders, identify:

• Staples — items in 70%+ of orders → write to staples.md
• Brand loyalty — same brand picked 80%+ of the time → write to preferences.md
• Dislikes — categories or items never or rarely ordered despite being common → note in preferences.md
• Co-purchase patterns — items that appear together in 60%+ of orders → write to patterns.md
• Seasonal trends — items that appear only in certain months → write to patterns.md

3. File formats

preferences.md:

# Taste Preferences
Last updated: YYYY-MM-DD

## Liked Categories
- Category (specific items, brands)

## Disliked / Avoided
- Items never ordered despite being common staples

## Brand Preferences
| Category | Preferred Brand | Notes |
|----------|----------------|-------|
| Milk | Brand | Fat %, size |

## Dietary Restrictions
- List any detected or user-stated restrictions

staples.md:

# Staple Items
Last updated: YYYY-MM-DD

| Product | Rohlik ID | Typical Qty | Frequency | Category |
|---------|-----------|------------|-----------|----------|
| Item name | 12345 | N | every order / every other | Category |

When processing orders via mcp__rohlik__get_order_detail, extract the Rohlik product ID for each item and store it in staples.md. This allows the shopping planner to call mcp__rohlik__add_to_cart directly without re-searching. Update IDs if Rohlik changes them (detect by name match with different ID).

meals.md:

# Favorite Meals
Last updated: YYYY-MM-DD

| Meal | Key Ingredients | Times Ordered | Last Cooked | Source |
|------|----------------|---------------|-------------|--------|
| Pasta bolognese | pasta, canned tomatoes, ground beef, parmesan | 5 | 2026-03-15 | co-purchase pattern |

Build this from co-purchase patterns and user-stated meals. When the same ingredient combination appears across multiple orders, infer a meal. When the user explicitly mentions a meal ("we're making pizza tonight"), add it with source "user-stated". Track Last Cooked from the most recent order containing all key ingredients.

planning-feedback.md:

# Planning Session Feedback
Last updated: YYYY-MM-DD

Tracks repeated user corrections during shopping planning sessions.
Only promote to preferences.md when a pattern is clear (3+ occurrences).

## Removals (user removed from proposed cart)
| Item | Times Removed | Last Session | Promoted |
|------|--------------|--------------|----------|
| Yogurt natural | 3 | 2026-03-23 | yes → preferences.md |

## Additions (user added items not in staples)
| Item | Times Added | Last Session | Promoted |
|------|------------|--------------|----------|
| Sparkling water | 2 | 2026-03-20 | no |

## Quantity Adjustments
| Item | Direction | Times | Last Session |
|------|-----------|-------|--------------|
| Milk | increased (1→2) | 4 | 2026-03-23 |

patterns.md:

# Shopping Patterns
Last updated: YYYY-MM-DD

## Co-Purchase Patterns
- Item A + Item B + Item C (reason/meal)

## Seasonal Trends
- Summer: items
- Winter: items

## Order Frequency
- Average interval: N days
- Typical order size: N-M items
- Typical cost: N-M CZK

spending.md:

# Spending Patterns
Last updated: YYYY-MM-DD

## Summary
- Average order: N CZK
- Monthly average: N CZK
- Orders per month: N

## Recent Orders
| Date | Items | Total | Notes |
|------|-------|-------|-------|

4. Pre-delivery analysis (run before every order)

Before building a new grocery order, refresh memory with any unprocessed completed deliveries:

1. Call mcp__rohlik__get_order_history to find recent completed/delivered orders
2. Read delivery-log.md to identify which orders have already been processed (by order ID or date)
3. For each unprocessed completed order, call mcp__rohlik__get_order_detail
4. Merge new data with existing consolidation — update preferences.md, staples.md, patterns.md, spending.md
5. Append new orders to delivery-log.md
6. Update fridge-state.md with delivered items (use actual delivery date, not order date)

This is an incremental process: existing consolidated data + newest deliveries → updated consolidation. No need to re-analyze the entire history each time.

5. Consumption model calibration

When processing completed orders, calculate actual consumption rates per product:

• For each staple item, measure the interval between consecutive orders containing it
• Compare against the default category rate from the fridge-tracker skill
• If the actual reorder interval consistently differs (3+ data points), write a calibrated rate to fridge-state.md under ## Calibrated Rates:

## Calibrated Rates
| Product | Category Default | Actual Avg | Data Points | Last Updated |
|---------|-----------------|------------|-------------|--------------|
| Milk 1.5% 1L | 5-7 days | 4 days | 6 | 2026-03-23 |
| Eggs 10pc | 14 days | 10 days | 4 | 2026-03-23 |

The fridge-tracker skill will use calibrated rates when available, falling back to category defaults.

6. Planning feedback promotion

After updating memory from completed orders, check planning-feedback.md:

• Removals with 3+ occurrences: Add item to preferences.md under "## Disliked / Avoided" and mark as Promoted: yes
• Additions with 3+ occurrences: Add item to staples.md (with Rohlik ID if known) and mark as Promoted: yes
• Quantity adjustments with 3+ consistent direction: Update the typical quantity in staples.md

This is the ONLY mechanism for learning from planning sessions — it requires repeated, consistent signals, not one-time corrections.

7. Waste detection

Compare reorder intervals against expected consumption:

• If a perishable item was ordered but NOT reordered within 2x its expected consumption period, it was likely wasted
• Note waste-prone items in preferences.md under a "## Waste Risk" section
• Consider reducing suggested quantities for these items

8. File size management

If any file exceeds 400 lines:

• Summarize older entries (keep last 3 months detailed, compress earlier data)
• For delivery-log.md: keep last 20 orders detailed, summarize earlier ones as monthly aggregates

Bootstrap Flow

If grocery/ directory is empty or doesn't exist:

1. Create the directory and all files
2. Fetch the last 10-20 orders from Rohlik
3. Process them to build initial profiles
4. Ask the user about household size, dietary restrictions, budget
5. Save to household.md
6. Confirm: summarize what you learned about their preferences