Architecture and data flow

2 min read

Remy is built as a set of small services on top of the bank's AI platform. It uses four platform primitives and introduces no new database: the knowledge base (storage, permissions, versioning, search, citation), the agent runtime (skills, tools, subagents), the MCP hub (connectors to Outlook, Teams, CRM, web search and any further system) and scheduled tasks (the nightly runs). Every run executes as the invoking user, with that user's permissions. Nothing leaves the tenant.

Four layers

Layer

What it does

Platform components

Sources

Client document folders in the knowledge base · Outlook, Teams, CRM and portfolio systems, web search read at run time · what the relationship manager / financial advisor tells Remy in chat · any further system connected as an MCP server. Matched once, then read by a unique ID - never by name.

Knowledge base, MCP hub, MCP servers

Client memory

Topic pages per client, live registers of open actions and meetings, an append-only change log, one prepared letter per open action. Every fact carries source, date, basis and confidence. Inherits the knowledge base's permissions, versioning, retention and backup.

Per-client folder in the knowledge base (markdown)

Remy

Aggregates all of a user's client memories every morning into one view. Renders a dashboard, meeting briefs, prepared drafts and chat with citations. Runs on schedule: mine · monitor · brief · dashboard. Drafts only.

Agent runtime, skills, scheduled tasks, side panel

Relationship manager / financial advisor

Reviews and sends from their own mailbox. Corrects and confirms in chat or on the page. Mail, calls and meetings re-enter through the sources on the next run.

Chat, dashboard, own mailbox

Client Memory and Remy

  • The memory builder collects. The only component that touches a source system. It reads what the connected systems know about a client and writes markdown, one folder per client, one file per topic. It raises nothing, sends nothing and derives nothing. It works in a deployment that has never heard of Remy.

  • Remy derives and can write back. It reads the memory and projects it into its own stores and pages - clients, meetings, actions, drafts, focus, dashboard. Those derived stores are Remy's own.

The test for any new piece of work: does it bring something in, or work something out? In goes to the memory builder, out goes to Remy. Two narrow write-backs into the client folder are documented exceptions, each with a single named owner under the one-writer-per-path rule: the live registers actions.md and meetings.md are written by the monitor skills and the meeting-brief skill, and a fact the relationship manager / financial advisor confirms in chat is written to the topic page that owns it. Nothing else derived re-enters the memory.

How Remy reaches a deployment

Remy is delivered with the contracted engagement; the hosting model - public cloud, private cloud, on premises - sets which models and outbound connections are available, and therefore which web-dependent skills can run.

Last updated