Proto-Mind

Inspiration

Most assistants either forget the user or hide how memory, learning, and action work. Proto-Mind explores a different model: a local-first cognitive operating system where continuity, evidence, goals, memory, skills, and permissions remain visible and operator-controlled.

What it does

Proto-Mind preserves useful context across sessions, separates remembered evidence from inference, and can explain why it believes something. A normal cognitive turn can, with exact process-session consent, become a bounded sequence of typed Experience events. The operator can inspect an Observe -> Interpret -> Recall -> Respond -> Reflect -> Verify episode, derive learning candidates, and promote lessons or skills only through separate supervised gates.

Its action layer is deliberately narrow. The active runner exposes four fixed read-only internal capabilities. Every run requires an exact command-specific confirmation and is checked against deterministic Registry and Policy metadata. Shell commands, arbitrary dispatch, background execution, network actions, and persistent approval are refused.

How we built it

Proto-Mind is written in Python 3.11 with a PySide6 Cognitive Control Room, tkinter fallback, local Ollama integration, and a deterministic mock backend. Local JSON/JSONL stores use atomic writes where mutation is allowed. Standard-library hashing, redaction, provenance, diagnostics, and receipts make behavior inspectable.

The macOS Demo Runway presents the full architecture as a guided twelve-step story: continuity, explicit Experience consent, explainable evidence, supervised learning, governance, and bounded action. The two sensitive buttons remain locked until their preceding previews generate exact process-bound commands.

Built during OpenAI Build Week

Proto-Mind existed before the event as a local memory, command, and safety prototype. During Build Week, Codex with GPT-5.6 helped extend it with bilingual cognitive continuity, pure retrieval and explicit telemetry, typed Experience provenance, exact consent, an explainable cognitive episode, supervised memory and skill lifecycles, bounded read-only action, the contest Showcase, and the Demo Runway.

The operator chose product direction, privacy boundaries, and autonomy limits. Codex inspected the architecture, proposed bounded milestones, implemented code and regression tests, diagnosed failures, verified protected-store hashes, and maintained a reproducible provenance trail. The public repository includes timestamped baseline/current manifests that distinguish prior work from the 78 relevant files and 480 tests added during Build Week.

Challenges

The hardest problem was not adding more commands. It was preventing hidden authority: retrieval must not silently mutate memory, capture must require exact consent, learning must not auto-promote, and execution must fail closed when metadata drifts. We repeatedly turned broad agent ideas into explicit contracts, receipts, doctors, refusal matrices, and regression tests.

Accomplishments and lessons

The submission contains 1,144 passing unit tests, 387 registered command prefixes across 41 categories, an inspectable supervised learning path, and a four-command read-only runner allowlist. The key lesson was that useful personal AI needs provenance and reversible gates as much as it needs model capability.

What's next

Next steps include richer semantic and episodic memory, verified multi-step supervised tasks, stronger relevance ranking, benchmark-driven reflection, and controlled skill reuse. Proto-Mind does not claim consciousness, unrestricted autonomy, neural self-training, or production-ready security.

Built With

• codex
• gpt-5.6
• ollama
• openai
• pyside6
• python