Architecture

Core Components

• API: persistence + orchestration of Projects and Tasks.
• Queue: Redis + RQ with three queues: high/default/low chosen by priority.
• Worker: executes agent jobs, writes results back via RQ result or pushes to API in future versions.
• FS Sandbox: agents can only read/write within FS_SANDBOX_ROOT if task allow_fs=True.
• Human tasks: claimed and completed via API; web UI (stub) will list and allow action.
• Knowledge Brain: Unified memory system with hybrid search, conversation memory, and decision tracking. See KNOWLEDGE_BRAIN.md for full documentation.

Knowledge Brain v2.0

The Knowledge Brain is Zenpower's state-of-the-art unified knowledge system:

Component	Technology	Purpose
Vector Search	pgvector + HNSW	Semantic similarity
Full-Text Search	PostgreSQL FTS + GIN	Keyword matching
Graph Search	Entity relationships	Multi-hop reasoning
Memory Hierarchy	MemGPT-style 4-tier	Infinite context
Reranking	ColBERT late-interaction	Precision improvement
Summarization	RAPTOR trees	Multi-level abstraction

Performance Targets:

• 100+ QPS, <50ms p95 latency
• 20% accuracy improvement on complex queries
• Military-grade access control and audit trails

Data model (MVP)

• Project(id, name, description, created_at)
• Task(id, project_id, title, type, status, priority, assignee, payload, result, depends_on[], allow_fs)

Extensibility

• Add worker job modules under apps/worker/jobs/*.py and reference the function name in payload.job (e.g., "agent_echo").
• Add auth, RBAC, audit logs for production.