The Self-Evolving
Multi-Agent Platform
Neurova is a powerful intelligent agent system supporting agent self-evolution, meta-cognition, multi-user management, multi-agent collaboration, intelligent task orchestration and multi-channel integration — helping you build enterprise-grade AI applications with neural perception.
Core Features
Powerful yet straightforward feature design, built to empower your AI applications
CogArch 2.0 Cognitive Architecture
A brain-partition model — cerebral cortex, cerebellum, brainstem and spinal cord — closing the full loop of observe → recall → reason → act → reflect.
Independent Brains · Shared Body
Each agent keeps its own memory and personality while sharing orchestration and execution, with hot reload and lazy loading.
Agent Growth System
Personality evolution, motivation drives, reflective learning and meta-cognitive optimization give agents the ability to keep growing on their own.
Three-Layer Agent Firewall
L0 entry gateway masking, L1 user-agent isolation and L2 file protection form a production-grade security stack.
Human-like Memory
Five layers — working, short-term, long-term, emotional and knowledge graph — with temperature decay curves and automatic memory consolidation.
Skill Ecosystem
Built-in skills, a public skill pool and an evolution engine in a marketplace model, with skills improving continuously through use.
Multi-Channel Communication
Unified access over REST, WebSocket, Telegram, WeChat, Discord and more, with real-time SSE streaming output.
Sleep and Wake Cycles
A Beta/Alpha/Theta/Gamma brainwave model of human sleep: automatic hibernation, memory sorting and scheduled wake-up.
Multiple Collaboration Modes
Sequential, parallel, master-slave and consensus modes over shared memory spaces, with recorded discussion decisions.
Hot-Pluggable Plugins
Automatic discovery, dependency resolution and hot load/unload, with configuration UI injected into the frontend — no restart needed.
A Complete API Surface
15+ modules of RESTful API covering cognition, memory, skills, growth, firewall and everything else.
A Biomimetic Intelligence Engine
Not a chatbot — an autonomous agent framework with a complete closed-loop cognitive architecture.
Latest Releases
See what Neurova has been shipping lately
Third public beta: full CUA desktop, self-evolving skills, three waves of security and resource repairs Beta
New Features
- Full CUA desktop capability: a UIA semantic desktop layer (control tree, semantic clicks, DPI multi-monitor coordinate chain, screenshot feedback per action) plus SOM visual numbered clicks, four run-permission levels (full/sandbox/review/auto), attach-mode authorization with action auditing, and two-tier MCP export
- Remote session plane: guest agent, session pool and Sandbox/RDP/SSH providers, with Linux/macOS driven over the SSH command surface and multi-host credential management
- New tool family: git, file_parse, deep_research, searxng, a conversational shell with stdin across turns, and one-click install of whitelisted MCP toolkits
- Self-evolving skill loop: an evaluation framework (three-tier eval sets, holdout splits, reflective mutation, benchmark gate) plus a trust funnel with lifecycle management and semantic-profile recall
- Growth and reflection loop: a motivation ledger with event-stream replay persistence, proactive-question feedback, and vote-to-reflection adjudication; sleep-phase cycles made real
- Channel ecosystem: end-to-end WeChat iLink QR pairing, per-agent multi-instance channel isolation, and platform message_id feedback
- Visuals: Galaxy starfield background, GlassSurface glass layer, and the liquid-glass treatment finalized across navigation and the input area
Bug Fixes
- Three waves through a 112-defect audit plus security closeout: router-level authorization on 37 endpoints, path traversal in generation, workspace IDOR, cross-session cancellation DoS, shell line-continuation bypass, SSRF, and full removal of the document surface
- A 37-item resource sweep: streaming concurrency slot leaks (throttled after 8 streams), whole-site freeze on iLink save, atomic persistence for the knowledge base, memory loss from WAL races, and blob leaks
- Dependency CVEs: axios 1.19, dompurify 3.4.12, echarts 6, postcss and nanoid upgraded across the chain
- Startup and stability: core startup cut from 48s to 1.5s, torch probe cached, Feishu connection made non-blocking, and single-flight refresh on 401
Improvements
- Unified scheduler event loop (B7): submissions go through a shared loop, so retries no longer exhaust workers
- Per-item session events with timeline persistence, a steering injection ladder, /review sub-sessions, and Claude Code-compatible hooks
- Settings center now persists for real (max_output_tokens, 429 retries, multi-host SSH credentials)
- Roughly 700+ new regression-prevention cases this week; the installer carries every beta3 change (built 2026-09-16 14:44)
Second public beta: native thinking channel, artifact preview dock, plan mode Beta
New Features
- Native thinking channel: Anthropic thinking blocks and Gemini thought parts are normalized at the protocol level into a single reasoning_content, so the reasoning timeline works across all models
- Artifact preview dock: a multi-tab dock on the right of the chat page registers tool output as artifacts automatically and previews them (image, audio, Markdown, HTML, text) without losing them on refresh
- Plan mode (/plan): requirements are gathered question by question, a plan is generated, and after approval it runs through the original chat chain
- Chat page rebuilt: a stepwise timeline that groups reasoning and tool calls in arrival order, a message queue card pinned at the top, KaTeX math, and a two-level model switcher
- Speech synthesis with moss-nano: the inference pipeline is wired end to end, on-demand loading saves 3.4GB of resident memory, and long text is no longer length-capped
- A full system architecture diagram (59 nodes across 6 regions, panoramic edition) ships in docs/ with this release
Bug Fixes
- Root-cause fix for file_operation scattering relative paths outside the sandbox
- Three fixes to the knowledge-graph extraction chain, including the reason it was always empty in production
- Chat image path: oversized images that triggered 413 are normalized and downsampled, and the attachment file_id contract is aligned, fixing whole turns of images vanishing silently
- Memory-system repairs: a closed loop for experience crystallization, sleep consolidation that no longer doubles memory, and pending-review memory isolated per proposer
Improvements
- Startup performance: core startup down from roughly 48s to roughly 1.5s (on-demand TTS, MCP moved to the background, embedding cache persisted to disk)
- Vision turns route by model capability: when the current model cannot see, a capable model handles the image and the original model is restored at the end of the turn
- Prompt upgrades: dynamic injections are enveloped (prefix caching plus immunity phrasing to resist injection), tools carry four-part descriptions, and Tool Search gains budget governance
- Single-file installer: a WPF step-by-step wizard plus silent install (/S /D=<directory>), defaulting to D:\Program Files\Neurova
- Themes and overlays: a color-scheme contract for light and dark, and overlays now mount on body uniformly, which fixes clipped dropdowns
First beta release Beta
New Features
- Windows desktop app released
- Substantial gains in agent capability
- Real-time speech output during conversation
- High-capacity memory
- 17 memory dimensions
- On-demand context pool
- Emotion factors influence the conversation
- Infinite-canvas visual workflows
- Large skill marketplace, ready out of the box
Bug Fixes
- Stronger tool-calling capability
- Ratchet pruning for recursive evolution
Improvements
- Experience crystallization
- Knowledge graph
- Windows desktop app released
Documentation
Get up to speed quickly, then go as deep as you like