The trickle: a lightweight, local-first LLM agent infrastructure layer by ellmos-ai.
Rinnsal gives small autonomous-agent projects the boring infrastructure they usually need first: SQLite memory, task state, connector I/O, chain automation, and an optional Ollama runner. It is extracted from BACH, but intentionally stays compact: Python stdlib only, no external runtime dependencies, no hosted service.
- Agent memory without a platform -- facts, notes, lessons, sessions, and prompt-ready context in local SQLite
- Task state for long-running agent work -- priority, status, ownership, and next-task selection
- Connector gateway -- Telegram, Discord, and Home Assistant behind one small abstraction
- Chain automation -- sequential LLM-agent runs with handoff, loop/once modes, logs, and stop/reset controls
- Embeddable core -- use Rinnsal inside another Python app, CLI workflow, or local agent stack
Use it when BACH is too large, USMC is too small, and you want a minimal local Python layer for LLM agent infrastructure.
- Memory -- Cross-agent shared memory with SQLite (facts, working memory, lessons learned, sessions)
- Tasks -- Simple task management with priorities, status tracking, and agent assignment
- Connectors -- Channel abstraction for Telegram, Discord, Home Assistant
- Automation -- LLM agent chain orchestration ("Marble-Run": sequential agent chains with loops, handoff, shutdown conditions)
- Ollama -- Local LLM runner for Ollama REST API (qwen3, mistral, etc.)
- Zero dependencies -- Pure Python stdlib, no external packages required
- Python 3.10+
pip install -e .from rinnsal.memory import api
api.init(agent_id="my-agent")
api.fact("system", "os", "Windows 11")
api.note("Current task: implement feature")
api.lesson("UTF-8 bug", "cp1252 encoding", "PYTHONIOENCODING=utf-8", severity="high")
print(api.context()) # Compact context for LLM prompts
print(api.status()) # Statisticsfrom rinnsal.tasks import api as tasks
tasks.init(agent_id="my-agent")
tasks.add("Implement feature X", priority="high", description="Details here")
tasks.add("Fix encoding bug", priority="critical")
for t in tasks.list():
print(f"[{t['id']}] {t['title']} ({t['status']})")
tasks.activate(1) # Set to active
tasks.done(1) # Mark as done
print(tasks.next_task()) # Next open task by priorityfrom rinnsal.auto.ollama_runner import OllamaRunner
ollama = OllamaRunner(model="qwen3:4b", base_url="http://localhost:11434")
if ollama.health():
result = ollama.run("Explain this code")
print(result["output"])
result = ollama.chat([
{"role": "system", "content": "You are a helpful assistant."},
{"role": "user", "content": "Hello!"}
])
print(result["output"])
print(ollama.available_models()) # List installed modelsfrom rinnsal.connectors import load_connector
# Token via ENV: export RINNSAL_TELEGRAM_TOKEN=123:ABC...
tg = load_connector("telegram")
tg.connect()
tg.send_message("chat_id", "Hello from Rinnsal!")from rinnsal.auto.runner import ClaudeRunner
runner = ClaudeRunner(model="claude-sonnet-4-6")
result = runner.run("Analyze the code in src/")
print(result["output"])rinnsal status # Overall status
rinnsal version # Version
# Memory
rinnsal memory status # Memory statistics
rinnsal memory fact system os "Win 11" # Store fact
rinnsal memory facts --json # All facts as JSON
rinnsal memory note "My note" # Store note
rinnsal memory context # Generate LLM context
# Tasks
rinnsal task add "My task" -p high # Create task (critical/high/medium/low)
rinnsal task add "Bug fix" -d "Details" # With description
rinnsal task list # Open/active tasks
rinnsal task list --all # Including done/cancelled
rinnsal task list --json # JSON output
rinnsal task show 1 # Task details
rinnsal task done 1 # Mark as done
rinnsal task activate 1 # Set to active
rinnsal task cancel 1 # Cancel task
rinnsal task reopen 1 # Reopen done/cancelled
rinnsal task delete 1 # Delete permanently
rinnsal task count # Count by status
# Chains (Automation)
rinnsal chain list # List chains
rinnsal chain start my-project # Start chain
rinnsal chain status # Status of all chains
rinnsal chain stop my-project # Stop chain
rinnsal chain log my-project # Show log
rinnsal chain reset my-project # Reset chain
rinnsal chain create # Create interactively
# Connectors
rinnsal connect list # Available connectors
rinnsal connect test telegram # Connection test
rinnsal connect send telegram ID "Text" # Send message
# Pipe (single call)
rinnsal pipe "Explain this code" # Single LLM callRinnsal looks for configuration in this order:
./rinnsal.json(current directory)~/.rinnsal/config.json(user home)
Secrets are set only via environment variables:
export RINNSAL_TELEGRAM_TOKEN="123456:ABC-DEF..."
export RINNSAL_DISCORD_TOKEN="..."
export RINNSAL_HA_TOKEN="..."Example config: config/rinnsal.example.json.
rinnsal/
|-- memory/ # SQLite-based cross-agent memory (from USMC)
|-- tasks/ # Task management with SQLite (priorities, status)
|-- connectors/ # Messaging channel abstraction (from BACH)
|-- auto/ # LLM agent chain orchestration + OllamaRunner
`-- shared/ # Config loader, event bus
Memory <-> Auto: Chains read/write context from Memory (opt-in)
Connectors <-> Auto: Telegram notifications via Connector
Event Bus: Decoupling layer between components
Rinnsal belongs to the ellmos infrastructure family:
| Project | Role |
|---|---|
| USMC | Shared SQLite memory primitive |
| Rinnsal | Memory + tasks + connectors + compact chain automation |
| MarbleRun / llmauto | Dedicated autonomous chain runner |
| BACH | Full text-based operating system for LLM agents |
Rinnsal is closer to an embeddable infrastructure layer than a full assistant product. For orientation, here is the difference to OpenClaw:
| Rinnsal | OpenClaw | |
|---|---|---|
| Focus | Minimal infrastructure layer: memory, connectors, automation | Full AI assistant: messengers, native apps, voice, skill marketplace |
| Philosophy | Provide building blocks, let the LLM handle the rest | All-in-one ecosystem with community-driven extensions |
| Memory | Structured SQLite (facts, lessons, working memory, sessions) | Session/workspace-oriented assistant context |
| Connectors | Telegram, Discord, Home Assistant (same abstraction pattern, growing) | Broad platform coverage |
| Dependencies | Zero: pure Python stdlib | Node.js and npm ecosystem |
| License | MIT | MIT |
In short: Rinnsal and OpenClaw share the connector gateway pattern, but take opposite approaches to complexity. Rinnsal stays minimal with zero dependencies and structured memory; OpenClaw is a large assistant ecosystem.
ellmos Rinnsal, Rinnsal LLM agent infrastructure, local-first LLM agent memory, Python stdlib agent framework, SQLite memory for LLM agents, lightweight agent connector gateway, BACH Rinnsal agent stack
MIT -- Lukas Geiger
Dieses Projekt ist eine unentgeltliche Open-Source-Schenkung im Sinne der §§ 516 ff. BGB. Die Haftung des Urhebers ist gemäß § 521 BGB auf Vorsatz und grobe Fahrlässigkeit beschränkt. Ergänzend gilt der Haftungsausschluss der MIT-Lizenz.
Nutzung auf eigenes Risiko. Keine Wartungszusage, keine Verfügbarkeitsgarantie, keine Gewähr für Fehlerfreiheit oder Eignung für einen bestimmten Zweck.
This project is an unpaid open-source donation. Liability is limited to intent and gross negligence (§ 521 German Civil Code). Use at your own risk. No warranty, no maintenance guarantee, no fitness-for-purpose assumed.