NemoFold

Architecture

Inspiration

People whose work lives in folders face a bad trade-off: local tools preserve privacy but often lack strong language reasoning, while cloud assistants tend to require broad uploads, forget prior file state, and return claims without precise evidence. File-maintenance agents add another risk: they may silently exceed the user's intent.

What it does

NemoFold is a private, persistent document agent that separates durable local authority from optional model reasoning. It inventories explicitly approved folders, maintains persistent document memory, answers multiple questions with exact quotes and locations, detects changed or conflicting versions, creates deterministic bundles and reports, and journals every file action for preview, recovery, and undo.

Its eight integrated workflows are Smart Inbox, Storage Policies, Universal Bundle Export, Continuous Folder Digest, Evidence Analyst, Version Resolver, Report & Artifact Studio, and Platform Proof.

How we built it

NemoFold uses a local-first hexagonal architecture. Its browser console, CLI, and agent skill share the strict nemofold.job.v1 contract and one application service.

Original files, absolute paths, the SQLite FTS index, inventory, policies, action journal, run ledger, validators, and generated artifacts remain local. When language reasoning adds value, a policy and privacy gate creates a hashed, path-free package containing only approved pseudonymized chunks, artificial source IDs, questions, a response schema, an explicit Nemotron model, and a bounded budget.

The Nebius Token Factory adapter accepts only the approved Nebius HTTPS origin, blocks redirects, requires current user-supplied prices, enforces a conservative cost ceiling, records durable transfer attempts, and validates returned source IDs and exact quotations before cloud_proof: true can be accepted. External model output cannot directly authorize file changes.

NemoClaw integration is implemented as a managed, validated skill and job-package route. Declared runtime metadata is deliberately not treated as execution proof.

Challenges

The hardest part was preserving a useful agent experience without handing the model unrestricted file authority. That required separating deterministic local capabilities from model reasoning, making external packages independently auditable, and treating failure states honestly. The implementation also supports reversible file actions, resumable runs, exact citation verification, cross-platform behavior, and one shared contract across web, CLI, and skill interfaces.

Accomplishments

Eight workflows built on five shared cores. Persistent local evidence and document memory. Exact quote and source-location verification. Reversible actions with preview, journal, resume, recovery, and undo. Deterministic bundles and multi-format reports. Fail-closed Nebius Token Factory adapter and verifiable result-package contract. Local Captain Nemo web console using the same application service as the CLI. 132 automated tests and green hosted CI on Ubuntu/Python 3.11 and 3.12 plus Windows/Python 3.12. Public MIT-licensed repository created after the competition submission window opened.

What we learned

A durable personal agent does not need one model to own everything. Deterministic local state, explicit policies, narrowly scoped reasoning packages, and verifiable evidence form a stronger boundary. The central rule is: reasoning may travel; file authority stays local.

What's next

The next gated steps are a sanitized live Nebius/Nemotron acceptance run, captured NemoClaw/OpenShell runtime evidence, a public working demo, a final video, and bounded user acceptance testing. Until those exist, the project truthfully reports cloud_proof: false and does not claim live NemoClaw execution.

Testing

The public repository includes synthetic examples, setup instructions, the strict job schema, architecture documentation, an offline demonstration, automated tests, and the jury-demo plan. No private user documents are required to evaluate the local workflow.

Repository: https://github.com/ellmos-ai/NemoFold

Hosted CI: https://github.com/ellmos-ai/NemoFold/actions/runs/33313331087