Cortex

Replaceable cognition. Persistent evidence. Human authority.

Cortex is a local-first continuity runtime that connects AI models to durable memory, provenance, governed tools, and challengeable measurement.

Models supply transient cognition. Cortex preserves the surrounding evidence, memory, session continuity and measurement. The human or host supplies operational authority. Changing a model does not need to erase the system's history.

Cortex is an advanced research prototype—not proof of general competence or autonomous self-improvement.

AI sessions end. Context drifts. Assumptions lose their sources. Models change. Persistent systems can reinforce mistakes, and even their measurement machinery can fail. Cortex aims to preserve useful continuity without turning memory into authority, and to make mistaken interpretations correctable without erasing history.

                       HUMAN / HOST AUTHORITY
                         bounds operational effect
                                  │
                    ┌─────────────▼─────────────┐
                    │           CORTEX          │
                    │ memory · context · tools  │
                    │ evidence · measurement    │
                    │ provenance · continuity   │
                    └───────▲───────────┬───────┘
                            │           ▼
                      MODEL / AGENT   TOOLS / REPOSITORY

Explore the system map for code, tests and evidence.

• Stored ≠ projected ≠ usable ≠ authorized.
• Evidence ≠ authority; hash integrity ≠ semantic validity.
• Evaluator result ≠ ground truth.
• Historical observations stay intact; interpretations may evolve.
• Measurement cannot outrank its instrument or its observation path.
• Adaptation cannot increase its own authority. UNKNOWN remains UNKNOWN.

Attach to an existing project from that project's directory:

uvx --from "git+https://github.com/jacksonjp0311-gif/Cortex@main" cortex-attach .

Attach options and privacy boundary. For reproducible installations, use a reviewed commit rather than moving main.

Run the local interface from a development checkout:

python -m pip install -e .
cortex ui

Use an isolated environment; see development setup. The local UI uses the Cortex runtime. Provider credentials are host secrets, not memory or evidence. Provider setup · Secret storage.

The next gate is complete observation-path assurance, not more autonomy. Three concrete transduction defects now have bounded fixes; historical failed audits remain preserved. No new live experiment is commissioned by this update.

The original eight-call repair observation remains 5/8. Corrected-transport replay is post-hoc evidence, not a new live 8/8 result. Current status and exact evidence.

A proposal, passing test, model confidence, memory item or receipt cannot grant permission. Host-issued capabilities and explicit operator controls bound actions. Detached worktrees do not sandbox the host OS. Shadow research must not become production policy. Security · Governance map.

Documentation hub · Agent entry · Knowledge map · Evidence

The complete previous README—including mathematics, experiments, commands, chronology and visual assets—is preserved as an archive. It is historical reference, not the current operational guide.

Cortex has measured and tested bounded mechanisms. General model improvement, semantic transfer, cumulative self-organization, consciousness and unrestricted recursive self-improvement are not established.

Contributing · Report a vulnerability · MIT license