Inspiration

Most AI assistants have "amnesia" between sessions. You tell them you're allergic to peanuts on Monday, and by Friday, they've completely forgotten. Worse, when your preferences change, naive memory systems just pile contradictory facts on top of each other. We wanted to build the missing memory layer: an agent that truly gets to know you over time, seamlessly superseding outdated information and curating exactly the right context to provide highly personalized answers.

What it does

Mnemo is a memory-augmented agent that operates on a continuous, five-step loop for every interaction:

Remember: Extracts durable memories (preferences, constraints, facts) from your conversation.
Supersede: Identifies and transparently retires outdated memories when your preferences change, implementing true "timely forgetting".
Recall (Budgeted): Retrieves only the most critical memories that fit into a limited context window, ranked by relevance, importance, and recency.
Reply: Answers your queries accurately, grounded in the highly curated memory context.
Broadcast: Streams the entire process, including the live 3D memory graph, to a React frontend via WebSocket.

How we built it

AI/Reasoning Layer: We used Qwen on Alibaba Cloud (DashScope). qwen-turbo handles fast memory extraction on every turn, while the more powerful qwen-plus drives supersession judgments and generates agent replies.
Memory Store: We integrated Cognee to build a durable, hybrid graph and vector store that persists across sessions.
Backend: A robust FastAPI backend that manages the agent loop, Cognee interactions, and real-time WebSocket broadcasting.
Frontend: A modern React 18 + Vite single-page application. We leveraged React Three Fiber and d3-force-3d to render a stunning, live 3D visualization of the user's memory graph.

Challenges we ran into

The biggest challenge was creating a reliable "timely forgetting" mechanism. We realized that simply deleting old facts wasn't enough, as it eliminated transparency. Instead, we developed a "supersession" engine where a new memory replaces an old one (e.g., changing from "vegetarian" to "pescatarian"). The old memory is marked as outdated and explicitly excluded from future context windows, while remaining visible in the user's history for trust and transparency.

Accomplishments that we're proud of

Context-Budgeted Recall: We're incredibly proud of our context-ranking algorithm. Mnemo never dumps its whole memory store into the prompt. It scores every active memory and precisely fits the "critical few" into the context budget.
The Live 3D Graph: Watching the agent build a mental model of you in real-time, visualized as a force-directed graph in the browser, is a magical experience.
Graceful Degradation: The system is engineered to fall back gracefully to heuristics if cloud keys or the memory store are temporarily unavailable, ensuring the agent loop never fully breaks.

What we learned

Building Mnemo taught us that maintaining long-term memory for an AI isn't just about storage—it's about curation. Figuring out what not to include in the context window (through supersession and budgeted recall) is just as important as the retrieval mechanism itself.

What's next for Mnemo

We envision Mnemo becoming a multi-tenant memory layer that could be plugged into any application. The next immediate steps include adding true multi-user isolation with authentication, and extending the recall mechanism to utilize embedding similarities for even richer, semantic retrieval.