Agent Architecture
Autonomous specialists with their own memory, tools, and personality — coordinated by the Orchestrator on your SovereignNode.
Fundamentals
An AIMOS agent is an autonomous domain specialist that combines four core components:
Dedicated SQLite database with semantic, episodic, and procedural long-term memory.
Domain-specific tool collection: SQL queries, REST calls, file operations, domain-specific functions.
Local model via local LLM runtime. The agent builds the complete prompt from system prompt, memory context, and user query.
Communication channel to the user: Telegram, email, voice, or dashboard.
How It Works
An AIMOS agent does not work like a chatbot — it works like a case worker with its own desk, its own filing system, and its own daily routine.
Remembers facts, corrections, preferences — permanently. Semantic search finds memories even with different wording.
Writes summaries, research results, and conversation notes into its own workspace — like a case worker maintains their filing cabinet.
Appointments, deadlines, follow-ups. The agent sees at every interaction its due and overdue appointments — and acts proactively.
Open tasks, delegations, waiting lists. Automatically updated during the Dreaming cycle.
Once daily, the agent independently checks calendar, emails, and open tasks — even without anyone addressing it.
During quiet phases, the agent processes its conversations: extracts facts, updates notes and todo lists, consolidates its memory.
Agents communicate with each other — the finance agent asks the logistics agent for inventory values, without burdening the user.
For complex questions, the agent automatically escalates to a more powerful AI model — anonymized and audit-proof logged.
The result: An AIMOS agent has everything a case worker needs — memory, notes, calendar, todo list, colleagues, a daily routine, and the ability to get help when needed. The only difference: It works around the clock, never forgets anything, and is available in seconds.
SKILL_REGISTRY
Each agent receives exactly the skills it needs — modularly assembled from the central registry. Custom skills can be developed and added at any time — for any industry, any ERP system, any workflow.
FEM calculation, profile optimization, DXF processing, code compliance checks, BOM export.
Read, write, search, list files. Workspace management with public/private separation.
Web research via Brave Search API. Current information without cloud dependency.
Send and receive emails via IMAP/SMTP. HTML and attachments.
POP3 mailbox monitoring. Detects new emails and automatically triggers actions.
Whisper STT + Piper TTS. Speech recognition and synthesis in all languages, processed locally.
Inter-agent file exchange via public/ folder. Read, write, list.
SFTP access to remote machines via Tailscale VPN. Fetch and deposit files from workstations.
Scheduled tasks: daily routine, recurring checks, reminders.
Create, query, modify, delete appointments. Proactive reminders for due dates.
Contact management: people, companies, phone numbers, email addresses, notes.
Permanently store notes and key-value pairs. Agent-owned knowledge base.
Escalation to a stronger model. Anonymized request to cloud LLM for complex analyses.
Turkish holidays, Ramadan, Kandil nights. Cultural context awareness.
Football results and schedules. Proactive notifications for important matches.
ETA V8 accounting via Firebird/InterBase. Account queries, open invoices, daily overview.
ETA V8 accounting via Microsoft SQL Server. Same functions, different database backend.
Browser automation for web portals without API. Fill forms, extract data.
Connectivity
The agent communicates via connectors — standardized interfaces to users, systems, and other agents. New connectors are continuously developed and can be added at any time for your specific IT landscape.
Text, voice messages, documents. Proactive messages for reminders, alerts, and results. Shared Listener for all agents.
IMAP/SMTP for sending and receiving. POP3 monitoring for incoming mailboxes. HTML format and file attachments.
Whisper STT + Piper TTS — completely local. Speech recognition and synthesis in all languages, without cloud services.
File access to workstations via Tailscale VPN. Shared folder for DXF, PDF, Excel — encrypted and without open ports.
PostgreSQL, MSSQL, Firebird — exclusively SELECT queries. No write access to production data. Read-only by design.
Universal API connectivity for ERP, CRM, inventory management. GET, POST, PUT with configurable authentication.
3-Tier Memory
Three memory types, hybrid search, and a Dreaming cycle for consolidation.
During idle time, the agent analyzes its conversations with an LLM call, extracts facts, updates notes and todo lists, consolidates its memory, and creates weekly reports.
Like the human brain during sleep — the agent condenses experiences into knowledge, removes redundant entries, and strengthens important connections. The result: more precise answers with lower token consumption.
Like the human brain during sleep, AIMOS consolidates memories during idle time:
Token Optimization
Dynamic balance between memory depth and response length — within the limited context window.
| Level | Memory | Response |
|---|---|---|
| 1 | Minimal | Very long |
| 2 | Reduced | Long |
| 3 | Balanced | Medium |
| 4 | Deep | Short |
| 5 | Maximum | Minimal |
Coordination
Inter-agent messaging, shared files, and architecturally enforced data privacy.
Agents of the same connector family share a communication channel. Messages to the family are automatically routed to the responsible agent.
Maximum 3 back-and-forth rounds between two agents per request. Prevents endless loops and uncontrolled token consumption.
Each agent has its own workspace — like a locked desk. Only deliberately shared files are visible to others.
Reports, exports, BOMs — everything other agents need. The agent decides what to share.
Notes, calendar, contacts, todo lists, profile databases. No other agent has access — architecturally enforced, not by policy.
Example: The engineering agent reads a DXF file from the engineer's workstation, creates a BOM, and places it in public/. The finance agent reads the BOM and calculates material costs — without access to the engineering notes.
Security
Three trust levels determine which tools an agent may use. Like an operating system kernel.
Lowest risk. The agent may only read: recall memory, view configuration. No access to external systems, no write access.
Read access to external systems. SQL SELECT, REST GET, file read. No write access to production data. Standard ring for most agents.
Full access including write operations. Requires explicit authorization by the administrator. Human-in-the-loop for critical actions.
Practical Examples
Three preconfigured agent types for typical industrial environments.
Ring 1 — Standard
Calculates structures with integrated FEM engine (PyNite), optimizes steel profiles by weight and cost, checks seismic codes (TBDY 2018, Eurocode). Creates BOMs as drafts — the engineer reviews and approves.
Ring 1 — Standard
Connects Uyumsoft, SAP, Navision, and other ERP systems. Monitors inventory levels in real-time, reports critical shortfalls, and automates purchase requisitions via SQL and REST connectors.
Ring 1 — Standard (Schreiben: Ring 2)
Integrates ETA V8, DATEV, and other accounting systems. Analyzes cash flow, prepares bookings, and delivers up-to-date balances. Booking finalization only by authorized clerks.
These are examples. AIMOS agents are individually configured per company — each agent is tailored to the specific systems and requirements.