All case studies

An AI Agent That Remembers

Building a personal business support agent with persistent memory that turns scattered information into sourced, trustworthy answers.

Industry
AI, Business Operations

4

Source types unified: email, meetings, documents and business systems

Summary:

Many businesses struggle to pull together, on the spot, all the information they need for a specific goal. Client details, project decisions and commitments sit across emails, meetings, documents and business systems. Context is easily lost and becomes legacy no one looks at until it is needed. Managers and team members then spend precious time combing through data to find what they need.

We developed a personal business support agent with advanced memory to close that gap. It supports questions about clients and projects, reviews of commitments and blockers, meeting preparation and business summaries. Persistent memory preserves context between conversations and connects every answer to its original sources.

The Solution: Advanced Memory with Business Context

At the core of the agent, its memory is built for business use rather than simple chat history. It combines structured records for clients, projects, people and finance with searchable communications and documents.

  • Decisions and responsibilities: topics, decisions, responsibilities and unresolved questions are retained together with their source references.
  • Two ways to search: semantic and keyword search find relevant information, while exact retrieval checks the underlying evidence.
  • Context that stays attached: the memory preserves who said what, when it happened and whether it was a proposal, condition or confirmed decision.
  • Honest about gaps: conflicting or missing information remains visible for review rather than being smoothed over.

An Agent Guided by Tools, Skills and Manuals

Codex provides an interactive agent environment. A tools layer gives it access to business memory and connected services, while reusable skills describe how to carry out tasks such as reviewing current work or investigating project risks. Manuals define the evidence rules, data routes and procedures.

The agent selects the relevant guidance, retrieves supporting information and prepares an answer or artifact. Business actions follow the applicable authorization and review process.

Controlled AI Processing

Defined AI workflows support the agent by routing requests, extracting facts and commitments, processing meeting content and drafting reports. These functions run within software workflows with validated outputs and bounded inputs.

  • Citations preserved: extracted statements retain their context and citations.
  • Review built in: proposed actions remain subject to review.

The result combines flexible reasoning with repeatable processing, while keeping consequential business decisions under human control.

Technologies

The platform combines Python, FastAPI and typed tools with PostgreSQL for structured and accepted state, MongoDB for content, and Qdrant for semantic and lexical retrieval. It uses Codex, managed agent conversations and direct model calls for defined AI tasks. A TypeScript web portal extends access to project information and workflows.

What’s Next: The First Phase of a Business Digital Twin

The personal support agent and its advanced memory form the first phase of this ongoing digital twin project. The broader system brings together source ingestion, persistent business memory, tools, skills and manuals with project spaces, responsibilities, work, finance, reports and alerts.

Its purpose is to maintain an evidence-based view of the business and help people understand what has happened, what needs attention and who should act.

Share:

Next step

Let's talk about your project.

Tell us what you are trying to build. We will read your brief and set up a call to see whether we are the right fit.