Memory for financial services

You are building an app where an AI agent helps a client with their accounts, meeting notes and instructions. Each client has their own memory, and the agent reads only that client's memory. Your team decides what the agent may say. Actions the agent takes pass through the app's permission settings: act without asking, ask the user first, or never allowed.

• Facts change. A client's goals, risk preferences and contact details can change. Your app needs to know which value is current.
• Dates matter. A changed value with a newer timestamp becomes the current value. The earlier value stays answerable by date, so you can show what the client said at a given time.
• Sources need naming. When the agent recalls a fact, it should be able to point to where that fact came from. Each recalled item lists its sources, with source id, label and time.

• Separate memory for each client. Each person's memory is kept apart from every other person's.
• Dated facts. A new value with a newer timestamp becomes current, and recall can run as of an earlier date.
• A record of agent actions. Each action has a target, a status and a time, and you can export the record.
• Deletion on request. You can delete a client's data points, or erase a client's data entirely.

• Summarise a client's current goals and stated risk preferences before a review meeting.
• Draft a follow-up note from a meeting record, with the source of each point named.
• Read the current value of a contact detail, and the earlier value by date.
• Hand off a step that needs a person, such as a signature or a call, and be told when the owner marks it complete.

1. Have your coding agent follow https://harny.ai/auth.md to set up the app, or call the Memory API from your server with your app key.
2. Send a client's notes with POST /v1/memory/ingest, using a stable id and a timestamp for when each fact happened.
3. Check each send with GET /v1/memory/ingest/{ingestionId} until its status is completed.
4. Ask for a client's current facts with POST /v1/memory/recall, using the client's identity and, where needed, a queryTimestamp.