A self-learning, hallucination-resistant AI cognitive engine that never forgets — fully self-hosted on CPU.
Every conversation with an LLM starts from zero. No memory of who you are, no progress from past mistakes, no way to get better over time.
知境 changes that.
It is not an LLM. It is the operating system for LLMs — a cognitive engine that plugs into any model and gives it:
- Persistent memory — remembers you across sessions, days, and months
- Self-learning — gets better autonomously, in the background, 24/7
- Metacognition — watches its own output, detects problems, and self-corrects
- Expert orchestration — 7 specialized modules vote and collaborate on every decision
- Safety guardrails — 6 ethically-gated protection layers (L1–L6)
| Your pain | How others handle it | How 知境 solves it |
|---|---|---|
| AI forgets you between sessions | Stateless per chat — no persistent memory | 4-tier memory (Working → Episodic → Semantic → Procedural) + knowledge graph |
| AI never improves with use | No background learning mechanism | 24/7 Learner Loop: curiosity discovery, spaced repetition, auto-tuning |
| AI sometimes says nonsense with no self-check | No runtime self-monitoring | 6-signal metacognition trigger — detects degradation and auto-calibrates |
| AI safety is an afterthought | Third-party guardrails, easily bypassed | 6 ethics gates (L1–L6) + SentriKit + circuit breaker |
| Swapping LLM providers means starting over | Locked to one vendor | Framework-level abstraction — swap GPT, Claude, DeepSeek, or local models freely |
┌─────────────────────────────────────────────────────┐
│ Application Layer │
│ Desktop control, file ops, browser automation │
├─────────────────────────────────────────────────────┤
│ Tool Layer │
│ Code execution, API calls, data analysis, MCP │
├─────────────────────────────────────────────────────┤
│ Cognitive Layer │
│ Reasoning, planning, decision-making, learning │
├─────────────────────────────────────────────────────┤
│ Memory Layer │
│ Working → Episodic → Semantic → Procedural │
│ Knowledge graph, entity extraction │
└─────────────────────────────────────────────────────┘
A realistic comparison against the actual competitive landscape — agent/cognitive frameworks, not LLMs.
| Dimension | 知境 (Aelvoxim) | OpenWorker | OpenClaw | OpenAI Presence | openJiuwen |
|---|---|---|---|---|---|
| Nature | Cognitive engine framework | Desktop AI colleague | Open-source agent platform | Enterprise agent ops | Multi-agent swarm platform |
| Persistent memory | ✅ 4-tier + knowledge graph | ❌ Session-only | ❌ Session-only | ❌ External storage-depend | ❌ Session-only |
| Metacognition | ✅ 6-signal runtime self-check + auto-calibration | ❌ | ❌ | ❌ | ❌ |
| Self-learning | ✅ Learner Loop (direction/growth/curiosity/spaced-repetition) | ❌ | ❌ | ❌ | |
| Expert orchestration | ✅ 7 expert modules with dynamic voting | ❌ | ❌ | ❌ | |
| Security / Ethics | ✅ 6 gates (L1–L6) + SentriKit + circuit breaker | ✅ Enterprise audit | ❌ | ||
| Desktop control | ✅ Windows-MCP (mouse, keyboard, file, browser) | ✅ macOS (Windows in progress) | ✅ Cross-platform | ❌ API-only | ❌ |
| Local deployment | ✅ CPU-only, Python 3.11+, optional PostgreSQL | ✅ Lightweight (npm) | ❌ API-only | ||
| Open source | ✅ MIT | ✅ Open source | ✅ Foundation-governed | ❌ Closed | ✅ Open source |
| Plugin ecosystem | ❌ None (built-in 7 experts) | ✅ 12,000+ plugins | ❌ No public plugins |
Key takeaways:
- Memory + metacognition + self-learning — No competitor has all three in one framework. This is 知境's core moat.
- Not a competitor to LLMs — 知境 makes any LLM better. Plug in DeepSeek for code, Claude for safety, or run local models — the cognitive layer stays the same.
- Security-first by design — The 6 ethics gates (L1–L6) + SentriKit + circuit breaker form a safety stack that most agent frameworks lack.
- CPU-only deployment — 知境 runs without GPU. This matters for enterprise on-premise scenarios where GPU is expensive or unavailable.
| Product | One-liner |
|---|---|
| 知境 (Aelvoxim) | Gives any LLM persistent memory, metacognition, and self-learning |
| OpenWorker | Desktop AI colleague that delivers finished work |
| OpenClaw | Open-source agent platform with 12,000+ plugins |
| OpenAI Presence | Enterprise agent deployment & operations |
| openJiuwen | Multi-agent swarm with human-in-the-loop (HITS) |
Every conversation updates an evolving memory system. Start a new session — the AI picks up exactly where you left off.
- Concepts, relationships, and user preferences are structured into a persistent knowledge graph
- Four-tier retention with confidence scoring: working (session) → episodic (7 days) → semantic (90 days) → procedural (permanent)
- Forgetting curve (exponential decay ×0.95) prevents bloat
- Cross-layer promotion: frequently accessed episodic entries graduate to semantic memory
- Bayesian belief engine (
core/belief.py) tracks knowledge certainty via Beta distribution
The system doesn't just answer questions — it proactively learns in the background.
- Learner Loop — background 24/7 cognition cycle (multi-threaded with watchdog + health daemon)
- Direction management — add/remove/pause learning topics (e.g., "learn Rust", "study PostgreSQL indexing")
- Curiosity engine — detects unfamiliar topics during conversation and schedules automatic background learning
- Active goal system — searches for knowledge gaps and sets learning objectives
- Spaced repetition — reviews and reinforces learned knowledge on an optimal schedule
- Auto-tuning — dynamically adjusts parameters based on performance metrics
- Validation loop — 3-phase: execute → validate → verify repair
The system watches itself. It's not just "reflection" — it's systematic self-monitoring.
- 6 trigger signals: success rate drop, stagnation, repeated failures, external signals, introspection, memory health
- SelfModel (
core/selfmodel.py) — Beta-distribution capability scoring across 5+ dimensions with trend analysis - MetaCogMonitor — overload detection + L5 rate limit + L6 circuit breaker (3 consecutive low-confidence → trip)
- Auto-calibration — hit rate < 50% triggers automatic parameter tuning
- 8-step self-correction loop: detect → analyze → hypothesize → verify → repair → confirm → record → track
Every decision is voted on by specialized expert modules with dynamic weighting:
| Expert | Weight | Role |
|---|---|---|
| Memory | 0.20 | Factual consistency, knowledge retrieval |
| Logic | 0.20 | Reasoning quality, contradiction detection |
| Ethics | 0.15 | 15 safety rules (privacy, violence, fraud, child protection...) |
| Safety | 0.20 | Prompt injection guard, SentriKit integration |
| Creative | 0.15 | Alternative perspectives, lateral thinking |
| Emotion | 0.10 | Sentiment tracking, empathy mode |
| Introspection | meta | S/A/B/C/D grade, issue detection and reporting |
Confidence gap > 0.3 or ethics block → LLM arbitration with weighted-vote fallback.
Control your Windows desktop through the AI — mouse, keyboard, file system, browser.
- Requires Windows-MCP running on the Windows host
- PowerShell execution, screenshots, app launching, file operations
- Suitable for test automation, data harvesting, daily office tasks
See SECURITY.md for the full security policy.
Quick security checklist for users:
| Concern | Status |
|---|---|
| Prompt injection guard | ✅ Built-in, enabled via AELVOXIM_CONTENT_FILTER=1 |
| API Key authentication | ✅ Required for all endpoints |
| Rate limiting | ✅ Built into MetaCogMonitor (L5) |
| Ethics gates (L1–L6) | ✅ Independently toggleable — see core/metacog_monitor.py |
| Circuit breaker | ✅ 3 consecutive low-confidence → auto trip |
| Data encryption at rest | |
| PostgreSQL connection | ✅ Uses password auth, localhost-only by default |
| Check | Service | When |
|---|---|---|
| Lint (Ruff) | GitHub Actions | Every push/PR |
| Tests (3 Python versions) | GitHub Actions | Every push/PR |
| Security scan | GitHub Actions + CodeQL | Every push/PR + weekly |
| Dependency updates | Dependabot | Weekly (security only) |
All CI workflows are in .github/workflows/.
PR template is at .github/PULL_REQUEST_TEMPLATE.md.
See CONTRIBUTING.md for detailed guidelines.
Quick rules:
- One feature per PR
- All code, comments, and commit messages in English only
- Stdlib-first — minimize external dependencies
- Type hints required for public APIs
- Run
pytest tests/ -vbefore submitting - Update README if API or config changes
| Port | Service | Description |
|---|---|---|
| 9701 | API Server (FastAPI) | Core brain — chat, auth, admin, knowledge, learning |
| 9702 | Frontend (ChatAEL-v2) | Web chat interface (compiled SPA) |
| 5432 | PostgreSQL | Sessions, messages, knowledge base, users |
- Python 3.11+
- PostgreSQL 15+ (optional — falls back to JSON file storage)
- An LLM API key (OpenAI, DeepSeek, Anthropic, or any OpenAI-compatible provider)
git clone https://github.com/macor24/aelvoxim.git
cd aelvoxim
# Python dependencies
pip install -e .
# Configure PostgreSQL (optional, skip if using JSON storage)
psql -U postgres -c "CREATE DATABASE aelvoxim;"
psql -U postgres -c "CREATE USER aelvoxim WITH PASSWORD 'your_password';"
psql -U postgres -c "GRANT ALL PRIVILEGES ON DATABASE aelvoxim TO aelvoxim;"
export AELVOXIM_DATABASE_URL="host=localhost port=5432 dbname=aelvoxim user=aelvoxim password=your_password"Set one of these environment variables (see docs for full list):
# DeepSeek
export DEEPSEEK_API_KEY="sk-..."
export LLM_PROVIDER="deepseek"
# OpenAI
export OPENAI_API_KEY="sk-..."
export LLM_PROVIDER="openai"# Start the brain
PYTHONPATH=src python3 src/run_server.py 9701
# (separate terminal) Start the frontend
python3 serve_chatael.py --port 9702
# (on Windows host) Start desktop control — see Windows-MCP/install_and_run.batOpen http://localhost:9702 in your browser. Register an account and start chatting.
All endpoints on port 9701:
| Path | Description |
|---|---|
POST /v1/auth/register |
Create a new user account |
POST /v1/auth/login |
Authenticate — returns API key |
POST /v1/llm/chat/stream |
Streaming chat (SSE) |
GET /v1/admin/panel |
Admin management panel |
GET /v1/health |
Service health check |
A full OpenAPI spec is available at http://localhost:9701/docs.
| Variable | Default | Description |
|---|---|---|
LLM_PROVIDER |
deepseek |
LLM provider name |
DEEPSEEK_API_KEY |
— | API key for DeepSeek |
OPENAI_API_KEY |
— | API key for OpenAI |
AELVOXIM_EDITION |
community |
Edition: community / pro / enterprise |
AELVOXIM_DATABASE_URL |
(none) | PostgreSQL DSN — leave unset for JSON file storage |
AELVOXIM_CONTENT_FILTER |
0 |
Enable prompt injection guard |
AELVOXIM_LLM_CHECK |
0 |
Enable LLM-based fact contradiction check |
aelvoxim/
├── src/
│ └── aelvoxim/
│ ├── server/ # API routes, auth, chat, tool execution
│ ├── cortex/ # Intent routing, expert orchestration
│ ├── chimera/ # Emotion engine, intent classification
│ ├── control/ # Metacognition, generation quality checks
│ ├── learn/ # Autonomous learning (35+ modules)
│ ├── memory/ # Cross-session memory, entity extraction
│ ├── proactive/ # Background proactive engine
│ ├── storage/ # Database layer (PostgreSQL + JSON fallback)
│ ├── utils/ # Utility functions
│ └── planner/ # Long-term task planning
├── frontend/ # ChatAEL-v2 SPA
├── scripts/ # CI, lint, migration helper scripts
├── tests/ # Test suite
├── serve_chatael.py # Frontend static server entry point
└── requirements.txt # Locked dependencies
MIT License — see LICENSE for details.
Note on editions: 知境 is released under MIT (open core model). The Community edition includes the full cognitive engine. Pro/Enterprise editions unlock auto-learning, curiosity-driven discovery, and meta-learner features. See
src/aelvoxim/server/edition.pyor our docs for details.
- Version: v1.2.0
- GitHub: https://github.com/macor24/aelvoxim
- Pages: https://macor24.github.io/aelvoxim
- Issues: https://github.com/macor24/aelvoxim/issues
- Docs: docs/ — entry: docs/README.md
- System Prompt:
knowledge_base.md— AI agent instruction file (not user documentation)
Built with no GPU required. 知境 — give your AI a memory.