EvoCortex: Frontier Intelligence & Hyper-Automation Suite

EvoCortex is a unified research and engineering monorepo combining cutting-edge agent foundations with concrete hyper-automation engines.

It is designed to solve the two biggest blockers of modern AI systems:

1. The Cost & Fragility Barrier: Automations break when websites change CSS, and agent tokens cost a fortune.
2. The Intelligence Bottleneck: Agents remain static, amnesic, and blindly execute dangerous actions without mental simulation.

---

The 5 Unified Engines

flowchart TD
    subgraph Automation ["⚡ Practical Hyper-Automation Layer"]
        GhostBridge["1. GhostBridge<br/>(Zero-Cost OpenAI API Gateway via Flat-Rate Browser)"]
        Unbrowse["2. Unbrowse<br/>(API-First Web Automation via Network Sniffing)"]
    end

    subgraph Intelligence ["🧠 Frontier Cognitive Agent Layer"]
        GhostWorld["3. GhostWorld<br/>(Active Inference & OS World Model Simulation)"]
        SurpriseMem["4. SurpriseMem<br/>(Google Titans Neural Test-Time Memory)"]
        EvoHarness["5. EvoHarness<br/>(Recursive Self-Evolution & Tool Synthesizer)"]
    end

    GhostBridge -->|"Unlimited Free Tokens"| EvoHarness
    GhostBridge -->|"Zero-Cost Inference"| GhostWorld
    Unbrowse -->|"Clean JSON Data Stream"| GhostWorld
    GhostWorld -->|"Simulated Trajectories"| EvoHarness
    EvoHarness <--> SurpriseMem

Loading

---

1. 🌐 unbrowse (packages/unbrowse)

Web automation that never breaks.

• The problem: Traditional Playwright/Selenium automations click CSS buttons (.btn-primary). When websites update their HTML/CSS or throw Cloudflare CAPTCHAs, scripts fail.
• The solution: Unbrowse runs in the background, intercepts the site's hidden private JSON fetch()/XHR traffic, extracts parameters and authentication tokens, and auto-generates standalone Python HTTP clients.
• Result: 100x faster execution, 0 browser RAM consumed, zero CSS selector maintenance.
• CLI:

  unbrowse parse network_log.har --output client.py

---

2. 🔌 ghost-bridge (packages/ghost-bridge)

Unlimited agent tokens for $0.

• The problem: Running autonomous coding agents (Claude Code, Cursor, Cline, AutoGPT) incurs heavy API fees ($20–$50/day), while developers already pay flat-rate subscriptions (ChatGPT Plus, Claude Pro).
• The solution: A zero-dependency local proxy exposing a standard http://localhost:8080/v1/chat/completions endpoint (OpenAI SDK compatible). It bridges existing browser sessions via Chrome DevTools Protocol (CDP) with full Server-Sent Events (SSE) streaming.
• CLI:

  ghost-bridge --port 8080

---

3. 🧬 evo-harness (packages/evo-harness)

Recursive self-evolution & tool synthesis.

• Grounding: EvolveR (2025), PILOT (2026), LATM (Large Language Models as Tool Makers)
• Analyzes successful multi-step agent trajectories and compiles them into permanent, typed, deterministic Python/WASM tools with auto-generated unit tests.
• Protected by PRM-Gate (Process Reward Model & AST invariant verifier) to eliminate regressions and misevolution.
• Future identical tasks run in 0 ms for 0 tokens.

---

4. 🧠 surprise-mem (packages/surprise-mem)

Neural test-time memory & dynamic context gating.

• Grounding: Google Research's "Titans: Learning to Memorize at Test Time" (2024/2025)
• Replaces naive flat vector RAG with neural associative memory.
• Uses surprise-gradient updates: predictable data decays naturally, while high-surprise observations update a persistent memory matrix with momentum.
• Includes ContextPrune to compress 128k prompt contexts down to high-density active tokens without losing needles in the haystack.

---

5. 🛡️ ghost-world (packages/ghost-world)

Active inference & OS state simulator.

• Grounding: Active Inference (Karl Friston) & Joint-Embedding Predictive Architectures (JEPA)
• Maintains a virtual shadow state of the OS, filesystem, and AST.
• Projects the exact file diffs and side-effects of shell/code mutations before executing them on the host system.
• Minimizes Expected Free Energy (EFE), deploying Monte Carlo Tree Search (MCTS) test-time compute when uncertainty is high.

---

Monorepo Layout

evocortex/
├── packages/
│   ├── unbrowse/           # Engine 1: API-first web automation generator
│   ├── ghost-bridge/       # Engine 2: Zero-cost local OpenAI API gateway
│   ├── evo-harness/        # Engine 3: Trajectory distillation & tool synthesizer
│   ├── surprise-mem/       # Engine 4: Titans neural test-time memory
│   └── ghost-world/        # Engine 5: Active inference OS simulator
├── docs/
│   ├── ARCHITECTURE.md     # Mathematical foundations and system design
│   └── ROADMAP.md          # Release schedule and modular extraction guide
├── tests/                  # Cross-engine integration tests
└── pyproject.toml          # Workspace configuration

---

Modular Independence

While all five packages live together in this monorepo for seamless interoperability, each package inside packages/ is strictly decoupled with its own pyproject.toml, test suite, and dependencies. Any package can be extracted into an independent standalone repository at any time.

---

Author & License

• Author: Junior Diomande (diomandejunior14@hotmail.com)
• License: MIT