Architecture
SingularCore is built on a bifurcated architecture that physically separates reasoning from execution. This guarantees that if the AI hallucinates, your system does not crash.
System Overview
Self-Improvement Cycles
The Brain and The Muscle
The Cognitive Engine (The Brain)
Written in asynchronous Python
The Brain handles orchestration, memory, and LLM routing. It never touches your raw filesystem directly.
- IntentClassifier & Task-Based MoA Router — Analyzes complexity, routes simple tasks to fast models, complex tasks trigger Mixture of Agents
- Unified Memory Tier — Four-layer architecture (Redis → PostgreSQL/pgvector → Kuzu Graph → File Archive)
- Skill Registry — Zero-trust plugin system with strict AST validation and blacklist pattern matching
The High-Speed Executor (The Muscle)
Written in memory-safe Rust and C++
The Muscle handles the dangerous work. It receives gRPC payloads from the Brain and executes them in isolated environments.
- SIMD-Accelerated Indexer (C++) — Crawls massive enterprise NAS drives and monorepos at near bare-metal speeds
- Firecracker Micro-VM Sandbox (Rust) — All AI-generated code runs in isolated sandboxes with strict cgroups resource limits
- Audit Ledger — Immutable, append-only SQLite WAL database for forensic logging
Self-Improvement Cycles
The Pulse
Every 60 seconds
Health checks, scheduled tasks, system monitoring. Your AI is always watching.
Dream Cycle
Nightly (2:00 AM)
Performance review, pattern extraction, self-optimization. Gets smarter while you sleep.
Micro-Dreams
Real-time
Instant repair on security breach or crash. Fixes itself in under 5 seconds.
Red Team Protocol
Twice daily
Self-penetration testing. Attacks itself to find and patch vulnerabilities.
Four-Tier Memory System
Redis — Hot Memory
Fast access for recent conversations and active context. Sub-millisecond retrieval.
PostgreSQL + pgvector — Episodic Memory
Semantic similarity searches across conversation history. Vector embeddings for context.
Kuzu Graph — Knowledge Graph
Relationships between codebases, concepts, and patterns. Structured knowledge retrieval.
File Archive — Cold Storage
Historical patterns and long-term logs. Compressed, indexed, rarely accessed.
Security Model: Tier-Based Actions
Actions are classified by impact level with corresponding approval requirements:
Tier 0: OBSERVE
Read-only operations. Automatic approval. Status, skills, dream summary.
Tier 1: COORDINATE
Sandboxed execution. Automatic. Goal submission, file reads, safe searches.
Tier 2: AUTONOMY
Local system changes. Internal review. File writes, staging deploys, code patches.
Tier 3: REASON
External/destructive actions. Human approval required. Production deploys, database changes.
Network Security
The agent has internet access, but it is physically restricted via Squid Proxy to a whitelist:
This resolves 90% of enterprise "rogue agent" fears. Even if the AI wanted to, it physically cannot reach your internal systems.
Blast Radius Control
All AI-generated code runs in Firecracker micro-VMs with strict resource limits:
| Risk | Docker | SingularCore (Podman + Rust) |
|---|---|---|
| Container Escape | Known vulnerability | Mitigated by design |
| Privilege Escalation | Requires root | Rootless by default |
| Blast Radius | Limited | Hardware-enforced isolation |