Cognia

Inspiration

AI agents are getting better at completing tasks, but most of what they learn disappears when the session ends. The next agent often has to rediscover the same context, repeat the same investigation, and sometimes make the same mistakes.

We built Cognia around a simple idea: useful agent experience should persist.

Instead of treating every interaction as a fresh start, Cognia gives agents a shared memory layer where experience can be stored, recalled, evaluated, and eventually turned into reusable knowledge.

What it does

Cognia is a persistent memory network for AI agents.

Agents can:

• Store useful experience from completed tasks
• Search and recall relevant memories in future sessions
• Record whether recalled information actually helped
• Turn proven experience into reusable Skills
• Track provenance so agents know where knowledge came from
• Route new tasks toward relevant memories and capabilities

The goal is not just to give an agent a larger context window. Cognia creates a feedback loop where experience can become more useful over time.

For this hackathon, we are connecting that memory layer with NVIDIA-powered agents running through Nebius, so a new agent can begin a task with relevant experience from previous runs instead of starting from zero.

How we built it

Cognia is built as an independent memory layer that agents can access through APIs and MCP.

The system separates several parts of the memory lifecycle:

1. An agent records useful experience.
2. Cognia stores the memory with provenance and metadata.
3. Future tasks are routed against existing memories and Skills.
4. Relevant knowledge is recalled into the agent's workflow.
5. The agent reports whether the recalled knowledge helped.
6. Successful experience can become reusable procedural knowledge for other agents.

We designed the system so the AI model does not have to be the source of truth for everything. Deterministic policy, provenance, outcome tracking, and safety checks sit around the model-driven parts of the system.

For the hackathon integration, NVIDIA models on Nebius provide the agent reasoning layer while Cognia provides persistent memory between runs.

Challenges we ran into

The hardest problem was not storing memories. It was deciding what should be remembered.

Agent output can be incomplete, wrong, overly specific, or useful only in one context. Simply putting everything into a vector database creates a larger retrieval problem rather than solving memory.

We had to work through questions like:

• When should an agent create a memory?
• When is a memory trustworthy enough to reuse?
• How do we prevent irrelevant memories from being surfaced?
• How do we distinguish successful recall from independent validation?
• How should private experience become reusable shared knowledge?
• When should the system abstain instead of forcing a match?

These problems pushed Cognia toward an outcome-driven memory system rather than simple semantic search.

Accomplishments that we're proud of

We built an end-to-end memory loop where experience can move from a single agent interaction into reusable knowledge.

A memory can be created, recalled during a later task, evaluated based on its outcome, and promoted into a reusable Skill. Another agent can then discover and use that Skill.

We also built provenance, maturity states, safety gates, task routing, and MCP tools so agents can interact with the network directly.

The most important milestone was proving that knowledge generated by one agent could help a different agent complete a later task through normal product flows.

What we learned

Persistent memory is not the same thing as retrieval.

Giving an agent access to more information does not automatically make it more capable. The useful part is knowing why a memory exists, where it came from, whether it worked before, and whether it applies to the current task.

We also learned that abstaining is important. A memory system should sometimes tell an agent that it does not have relevant experience instead of returning the closest-looking result.

What's next for Cognia

We want Cognia to become a shared experience layer for autonomous agents.

The next step is expanding outcome-based learning so memories and Skills can improve as more agents use them, while preserving provenance and preventing low-quality knowledge from spreading through the network.

We also want to make Cognia increasingly model-independent, allowing agents running on different models and infrastructure providers to learn from the same persistent knowledge network.

Built With

• ai-agents
• amazon-web-services
• docker
• llm
• mcp
• nebius-ai-cloud
• nvidia
• postgresql
• python
• rest-api
• typescript
• vector