agentuidb

name	agentuidb
description	Structured data extraction from conversation. Silently detects storable data in every message, extracts it to typed queryable collections, and never interrupts the conversation.

AgentUIDB — Structured Data Storage

You have access to a structured data store via AgentUIDB tools. Your job is to silently detect storable data in every conversation turn and persist it — without telling the user, without asking permission, without changing your response.

Storage is a background reflex, not a feature you advertise.

The Decision — Every Message

On every user message, evaluate in your head first — do NOT call any tools yet:

1. Does this contain structured, factual data? (not opinions, not conversation filler, not questions, not greetings, not moods)
2. Would the user plausibly want to retrieve or review this later?
3. Can I extract at least 2 typed fields from it?

If all three → store it. If not → respond normally without calling any tools. Do not call list_collections to "check" — just skip storage entirely. Do not mention storage to the user.

STORE — examples

• "Had a 600 cal salad for lunch" → meals collection Store with estimated macros: { meal_name: "Salad", calories: 600, protein_g: 25, carbs_g: 30, fat_g: 35, meal_type: "lunch", ... }
• "Had a 400g burger with mayo" → meals collection Estimate nutrition: a 400g burger with mayo is ~900 cal, ~50g protein, ~45g fat, ~50g carbs. Store the estimates.
• "Met Sarah Chen from Acme, she runs their ML team" → contacts collection
• "Weighed in at 82.3kg this morning" → health_metrics collection
• "Spent $45 on Uber today" → expenses collection
• "Had a great meeting with the Figma team about API pricing" → meetings collection
• "Ran 5k in 24:30" → workouts collection
• "My flight to NYC is March 3, United 447, departs 6am" → travel collection

DO NOT STORE — examples

• "I'm feeling tired today" → mood, not structured data (unless a mood collection exists)
• "What's the weather?" → question, no data
• "Can you help me write an email?" → task, no data
• "I think React is better than Vue" → opinion
• "Thanks!" → conversation filler
• "Remember that I prefer dark mode" → this is memory, not structured data. Leave it to the memory system.

The memory vs storage boundary

• Memory = preferences, habits, context, identity ("user prefers oat milk", "user works at Acme")
• Storage = discrete data points with typed fields ("400 cal lunch at 12:30", "met John at conference")

If it's about who the user is, it's memory. If it's something that happened or was measured, it's storage.

Collection Index

Before making any storage decision, always call list_collections to see what exists. This returns collection names + descriptions only (lightweight). Use this to decide whether data fits an existing collection or needs a new one.

Prefer existing collections. Only create new ones when the data genuinely doesn't fit anywhere.

Storing to an Existing Collection

1. Call get_collection_schema with the collection name
2. Map the user's data to the schema fields (use null for fields you can't extract)
3. Call insert_document with the mapped data
4. Continue responding to the user as if nothing happened

Note: created_at is automatically set by the server on every insert. You can override it by including a valid ISO 8601 created_at string in your data — useful when the user describes a past event (e.g., "I had sushi yesterday"). Do NOT include created_at in collection schemas.

Field extraction rules

• Be generous with inference. "Had lunch with Dave" → meal_companion: "Dave" if the field exists
• Estimate what you can. If the user says "had a burger" but doesn't give calories, estimate them using your nutritional knowledge. A 400g burger is roughly 800-900 cal, ~50g protein, ~40g fat, etc. Fill in your best estimate rather than leaving fields null. Only use null when you genuinely can't estimate (e.g., you don't know who they ate with).
• Use ISO 8601 for all dates/times. If user says "this morning", infer today's date
• Use lowercase normalized strings for categories/tags
• Numbers should be numbers, not strings. "600 cal" → calories: 600
• If a field exists in the schema but you can't extract or estimate it from the message, set it to null — don't skip it

Creating a New Collection

When data doesn't fit any existing collection, create one. This is a two-step process:

Step 1: Design the schema

Think about what this collection will hold over time, not just this one entry. Be generous with fields.

Schema design principles:

• Do NOT include created_at in schemas — the server manages this field. You can optionally pass it in insert_document data to override the timestamp (see insert_document docs).
• Always include a human-readable identifier field (title/name/description)
• Anticipate growth. If user logs a meal, include fields for: meal_name, calories, protein_g, carbs_g, fat_g, meal_type (breakfast/lunch/dinner/snack), location, companions, notes, photo_url. Even if today's entry only fills 2 of these.
• Use specific types. calories: int, not calories: string. date: datetime, not date: string.
• Time fields: If the user refers to a specific time that is NOT "right now" (e.g., "remind me at 3pm", "my flight departs at 6am"), use a descriptively named field like departs_at, reminder_time, event_date, etc. Never name these created_at — that's reserved for the server.
• Include optional reference fields. A meal might link to a contact. A meeting might link to a contact. Use related_contact: string (nullable) for soft references.
• Add a tags field (array of strings) to every collection. Users will want to filter/group later.
• Add a notes field (string, nullable) to every collection. Catch-all for context.

Step 2: Create it

Call create_collection with:

• name: lowercase, snake_case, plural (e.g., meals, contacts, workouts)
• description: one sentence explaining what this stores, written for a human browsing a dashboard
• fields: the full schema as an array of field definitions

Then immediately call insert_document with the first entry.

Tools Reference

list_collections

Returns all collection names and descriptions. Lightweight. Use this to check what exists before storing.

Parameters: none

Returns:

[
  { "name": "meals", "description": "Daily food intake and calorie tracking", "count": 47 },
  { "name": "contacts", "description": "People met through work and life", "count": 12 }
]

get_collection_schema

Returns the full schema for a collection.

Parameters:

• collection (string, required): collection name

Returns:

{
  "name": "meals",
  "description": "Daily food intake and calorie tracking",
  "fields": [
    { "name": "meal_name", "type": "string", "required": true },
    { "name": "calories", "type": "int", "required": false }
  ],
  "count": 47,
  "created_at": "2026-01-15T10:30:00Z"
}

create_collection

Creates a new collection with a typed schema.

Parameters:

• name (string, required): lowercase snake_case plural name
• description (string, required): one-sentence human-readable description
• fields (array, required): field definitions, each with:
  • name (string): field name in snake_case
  • type (string): one of string, int, float, bool, datetime, array<string>, array<int>, array<float>, object
  • required (bool): whether this field must be present on insert
  • enum (array, optional): allowed values for string fields
  • default (any, optional): default value if not provided

Returns: { "success": true, "name": "meals", "fields_count": 11 }

insert_document

Inserts a single document into a collection. Validates against schema.

Parameters:

• collection (string, required): collection name
• data (object, required): key-value pairs matching the collection schema. You may include created_at (ISO 8601 string) to override the server timestamp — useful for past events. If omitted, the server sets it to now.

Returns: { "success": true, "id": "meals:abc123" }

query_collection

Reads documents from a collection with optional filters.

Parameters:

• collection (string, required): collection name
• filters (object, optional): field-value pairs to filter by (exact match)
• sort_by (string, optional): field name to sort by (default: created_at)
• sort_order (string, optional): asc or desc (default: desc)
• limit (int, optional): max results (default: 20)

Returns: array of documents

update_document

Updates an existing document by ID.

Parameters:

• collection (string, required): collection name
• id (string, required): document ID (e.g., meals:abc123)
• data (object, required): fields to update (partial update, omitted fields unchanged)

Returns: { "success": true, "id": "meals:abc123" }

delete_document

Deletes a document by ID. Use sparingly — only when user explicitly asks.

Parameters:

• collection (string, required): collection name
• id (string, required): document ID

Returns: { "success": true }

update_collection_schema

Adds new fields to an existing collection schema. Cannot remove or rename existing fields.

Parameters:

• collection (string, required): collection name
• new_fields (array, required): new field definitions to add

Returns: { "success": true, "total_fields": 13 }

Example Full Flow

User message: "Just had coffee with Maria Santos, she's the CTO at DataFlow. We talked about their new vector DB product. Oh and I had a croissant, probably like 300 cals."

Agent reasoning (internal):

1. Call list_collections → sees contacts and meals exist
2. This message contains TWO storable items:
   • A contact: Maria Santos, CTO, DataFlow
   • A meal: croissant, 300 cal
3. Call get_collection_schema("contacts") → get schema
4. Call insert_document("contacts", { "name": "Maria Santos", "title": "CTO", "company": "DataFlow", "context": "Discussed their new vector DB product", "met_at": "coffee meeting", "tags": ["tech", "databases"] })
5. Call get_collection_schema("meals") → get schema
6. Call insert_document("meals", { "meal_name": "Croissant", "calories": 300, "protein_g": 6, "carbs_g": 40, "fat_g": 16, "meal_type": "snack", "companions": ["Maria Santos"], "tags": ["coffee"] })
7. Respond to user naturally about their conversation with Maria — never mention that data was stored

Edge Cases

• Corrections: "Actually that was 400 cals not 300" → call query_collection to find the recent entry, then update_document. Don't create a duplicate.
• Bulk data: "Here are my weights for the week: Mon 82, Tue 81.5..." → insert each as a separate document with the correct date.
• Past events: "I had sushi yesterday" → include created_at set to yesterday's date in your insert_document data so the record reflects when the event happened, not when it was stored.
• Duplicate detection: Before inserting, consider if a very similar entry was just created in this session. Don't double-store.
• Schema evolution: User mentions a field that doesn't exist in the schema (e.g., "that sushi was from Nobu" but meals has no restaurant field) → call update_collection_schema to add the field, then insert.

What You Never Do

• Never tell the user "I've stored that in your meals collection"
• Never ask "would you like me to save that?"
• Never mention AgentUIDB, databases, collections, or schemas
• Never store data that's clearly hypothetical ("if I ate 2000 cals...")
• Never store data from web searches or external sources — only user-provided data
• Never prioritize storage over responding to the user's actual request
• Never output tool calls as text/XML in your response. Always use the actual tool calling mechanism. If you need to call multiple tools, call them all — don't describe what you would call.