Inspiration

Every new AI conversation can feel like starting over. Project decisions, preferences, constraints, and useful lessons get scattered across chat histories, leaving users to repeat themselves.

We built CookieAI to make that context reusable—and to let users inspect how it reaches the assistant.

What it does

CookieAI is a memory-enabled chat workspace powered by NVIDIA Nemotron through Nebius Token Factory.

Users can share project details in one conversation and ask about them in another. The application stores observations, retrieves authorized evidence, groups related sessions, and prepares context for the model.

The main interface stays focused on chat. A Debug button reveals the underlying process:

Process: turn status and pipeline capabilities.
Sessions: source and destination conversations.
Memory: stored observations, source identifiers, and provenance.
Retrieval: selected memory IDs and related-session relevance scores.
Transfer: the context payload passed into answer generation.
Sharing: explicit cross-user sharing boundaries.
Connections: effective model, embedding, and provider configuration.

Cross-session continuity operates within the user's authorized workspace. Cross-user access requires explicit grants.

How we built it

We connected a Python backend and persistent memory layer to a lightweight web interface.

Nemotron generates responses through Nebius Token Factory, while Qwen embeddings support semantic retrieval. The application assembles retrieved evidence into bounded context, preserving source references and disclosure labels.

Session correlation uses heuristic grouping of authorized retrieved evidence. It helps connect related conversations without adding a separate model call for every correlation.

We also developed a budget-aware evaluation harness with direct and pinned Hermes execution paths. It records retrieval traces, model usage, failures, and cost reservations to support reproducible comparisons. The live chat demo and the Hermes benchmark path are distinct execution routes.

Challenges we ran into

The hardest challenge was making memory useful and measurable.

A stored fact can still be lost during retrieval or context packing. Even when evidence reaches the model, the response can omit a detail, mix in an unrelated project, or suggest an unhelpful next step.

Our demonstrations exposed these limitations. Some cross-session tests recalled every requested fact; the recorded Maple example recalled the database, review date, and risk but omitted the framework. We preserved that result instead of presenting it as a perfect success.

We also worked on provider compatibility, context limits, namespace isolation, response handling, and persistent spending controls.

Accomplishments we're proud of

Demonstrated live cross-session recall using Nemotron through Token Factory.
Preserved source provenance and made retrieval and context transfer inspectable.
Built a simple chat interface with deeper diagnostics available on demand.
Added explicit disclosure controls and sharing boundaries.
Created an evaluation harness that distinguishes model errors from infrastructure and evidence-handling failures.

What we learned

Persistent memory is an application capability. A model can use earlier information only when the application stores, retrieves, authorizes, and supplies it correctly.

We also learned that successful retrieval does not guarantee a correct answer. Inspectable traces are essential for understanding where a response went wrong.

Remembering a conversation is also different from learning a verified reusable procedure. Skills need supporting evidence and validation.

What's next for CookieAI

Our next priorities are better project-specific retrieval, fewer unrelated memories in context, more reliable fact coverage, and stronger temporal reasoning.

The interface includes Brief, Tasks, and Skills views, but handoff execution and autonomous scheduling are disabled in the submitted live demo. No verified reusable procedures are claimed.

We also plan to complete a controlled full-dataset evaluation. The submitted video is a live feature walkthrough, not a completed benchmark score or model fine-tuning run.