EvoMind: A Governed Cognitive Architecture for Persistent, Verifiable, Experience-Driven Agentic Autonomy

EvoMind is a local-first, governed cognitive architecture designed to separate cognitive capability, runtime authorization, execution, and verification into explicit system authorities rather than treating model output as direct permission to act.

This technical note documents EvoMind at a frozen first-party engineering snapshot dated 10 August 2026, corresponding to repository main commit 628de798bbc4ab1d729b584cd06231a7d372cf1e. The documented system integrates persistent state and memory, multi-step planning, fused desktop perception, governed tool and desktop execution, canonical skill lifecycle management, fresh post-action verification, semantic human demonstration learning, restart reconstruction, revocation, provenance, causal evidence, and experience-driven skill formation from ordinary verified successful execution.

The architecture implements a single durable skill-promotion authority in which discovery and evidence production do not independently grant executable authority. Candidate skills move through explicit evidence, shadow, approval, activation, and revocation states. Activated learned capabilities execute through existing governed runtime authorities rather than through unrestricted macro replay or an alternate execution engine.

The paper reports evidence including a 99-file perception/desktop reachability audit; complete disposition of the original 51 dormant capability donors; governed semantic desktop teaching without persistent pointer coordinates or raw typed-character trajectories; restart-surviving learned capabilities; causal and regression evidence; bounded experience-to-skill compounding; real Windows desktop validation; reusable cognitive-operator extraction; and project-reported MiniWoB++ progression from 261/625 (41.76%) to 407/625 (65.12%).

EvoMind is presented as an AGI-oriented cognitive architecture, not as established human-level AGI. This publication does not claim consciousness, sentience, human-level or superhuman general intelligence, unrestricted recursive self-improvement, autonomous source-code mutation, universal computer-use competence, distributed multi-node correctness, or independent third-party validation of the private implementation.

The deposit includes the technical whitepaper, citation metadata, bibliography, evidence manifest, package manifest, cryptographic checksums, rights notice, version metadata, and archival supporting materials.