MemoryOS

MemoryOS gives AI agents long-term, connected memory and FounderOS demonstrates it as a startup CEO’s second brain

Comment

Inspiration

AI agents are useful in a single conversation, but they often forget the people, decisions, constraints, and events that make their answers genuinely valuable over time.

MemoryOS was built to solve that problem: a persistent memory layer that lets AI systems remember, retrieve, and reason over meaningful context across many interactions and data sources.

For the hackathon, we built FounderOS on top of MemoryOS as a practical demonstration. FounderOS is not the core product—it is an example of what becomes possible when an AI agent has a company’s long-term memory.

What it does

founderOs is built to show capabilities of MemoryOS which is the main product . MemoryOS stores and organizes long-term company context , including f:

  • Customers and account history
  • Roadmap decisions
  • Investor meeting notes
  • Business metrics
  • Slack discussions
  • GitHub issues and engineering blockers

It makes that information retrievable when it matters. An AI agent can ask MemoryOS for relevant context and respond with an answer grounded in remembered company knowledge rather than only the current prompt.

FounderOS demonstrates this capability through a CEO-agent interface. A founder can ask:

  • Which customer needs my attention?
  • What could delay our enterprise launch?
  • What should the board update cover?
  • Which engineering blockers affect customer commitments?

Behind the interface, MemoryOS retrieves related memories across customers, roadmap items, discussions, metrics, and issues, giving the agent a connected view of the company.

How we built it

MemoryOS is the core system. It provides the memory ingestion, retrieval, context assembly, and persistence layer.

For the demo, we extended MemoryOS with:

  • Persistent memory retrieval for CEO questions
  • A context-grounded question-answering endpoint
  • Memory categories for customers, roadmap, investor meetings, metrics, Slack, and GitHub
  • LLM usage tracking and daily request limits
  • A graph-backed memory model using Neo4j
  • PostgreSQL-backed application data
  • A FounderOS frontend that visualizes what a MemoryOS-powered agent can do

The system uses FastAPI for the API, PostgreSQL and Neo4j for persistence and relationships, Docker for local infrastructure, and an OpenAI-compatible LLM API for natural-language reasoning.

Codex and GPT-5.6

We used Codex with GPT-5.6 as a development collaborator to build FounderOS on top of MemoryOS. Codex helped implement the CEO question endpoint, connect retrieval-grounded context to the LLM response flow, add durable daily LLM usage limits, configure the local frontend-to-API integration, troubleshoot Docker and database startup issues, and prepare the project documentation. GPT-5.6 and Codex were used during development for implementation, debugging, iteration, and integration; the repository contains the resulting API routes, migration, usage safeguards, and

Challenges we faced

The biggest challenge was ensuring the AI did not behave like a generic chatbot. A useful answer must come from remembered, relevant context.

We addressed this by making MemoryOS responsible for retrieving the relevant information before the LLM generates an answer. We also added request-budget controls because LLM calls are limited and should be spent on high-value reasoning rather than every interface interaction.

A second challenge was demonstrating a general-purpose memory system in a way that feels concrete. FounderOS became our use case: it shows how the exact same MemoryOS foundation could support sales agents, support agents, project managers, personal assistants, or any application that benefits from persistent context.

What we learned

The key insight is that an AI agent becomes significantly more useful when it can remember across time.

MemoryOS is designed as that memory foundation. FounderOS is simply one demonstration of the kinds of intelligent applications that can be built on top of it.

What's next

We plan to make MemoryOS easier to connect with external data sources such as Slack, GitHub, CRM systems, calendars, and documents. We also want to add source citations, access controls, richer graph exploration, and proactive memory-driven alerts.

The long-term goal is to make persistent, connected memory a reusable building block for every AI agent.

Built With

Share this project:

Updates