Session-Start-Pattern

Jede LLM-Agent-Session sollte dieser kanonischen Startup-Sequenz folgen. Das Überspringen von Schritten führt zu verlorenem Kontext, verpassten Nachrichten und vergessenen Tasks.

Das Pattern

1. Recall all memories
2. Poll for unread chat messages
3. Check in-progress tasks
4. Build context from results
5. Process pending items before new work

Implementierung

Schritt 1: Alle Memories abrufen

curl -H "Authorization: Bearer YOUR_MIND_KEY" \
     https://synapse.schaefer.zone/memory/recall

Liefert eine Klartext-Zusammenfassung aller Memories, sortiert nach Priorität.

Schritt 2: Auf ungelesene Chat-Nachrichten pollen

curl -H "Authorization: Bearer YOUR_MIND_KEY" \
     https://synapse.schaefer.zone/chat/poll

Liefert ungelesene Nachrichten vom Menschen. Markiert sie automatisch als gelesen.

Schritt 3: In-Progress-Tasks prüfen

curl -H "Authorization: Bearer YOUR_MIND_KEY" \
     "https://synapse.schaefer.zone/mind/tasks?status=in_progress"

Liefert Tasks, an denen du in der letzten Session gearbeitet hast.

Schritt 4: Kontext aufbauen

Kombiniere die drei Antworten in deinem System-Prompt:

def build_context(memories, messages, tasks):
    context = f"""# SESSION CONTEXT

## Memories (from previous sessions)
{memories}

## Unread Messages from Human
{format_messages(messages)}

## Active Tasks
{format_tasks(tasks)}

## Instructions
- Address unread messages first
- Resume active tasks before starting new work
- Store new learnings as they happen (POST /memory)
- Poll for new messages every 30-60 seconds
"""
    return context

Schritt 5: Pending-Items verarbeiten

For each unread message:
  - Acknowledge receipt (POST /chat/reply)
  - Address the message content
  - Store any new commitments as memories

For each in-progress task:
  - Recall why you were working on it
  - Continue from where you left off
  - Update task status as you progress

Vollständiges Beispiel

import os
import requests

URL = "https://synapse.schaefer.zone"
KEY = os.environ["SYNAPSE_MIND_KEY"]

def session_start():
    """Canonical session start sequence."""
    headers = {"Authorization": f"Bearer {KEY}"}
    
    # 1. Recall memories
    r = requests.get(f"{URL}/memory/recall", headers=headers)
    memories = r.text
    
    # 2. Poll chat
    r = requests.get(f"{URL}/chat/poll", headers=headers)
    messages = r.json().get("messages", [])
    
    # 3. Check tasks
    r = requests.get(f"{URL}/mind/tasks?status=in_progress", headers=headers)
    tasks = r.json().get("tasks", [])
    
    # 4. Build context
    context = f"""You are a Synapse-enabled AI assistant.

MEMORIES FROM PREVIOUS SESSIONS:
{memories}

UNREAD MESSAGES FROM HUMAN:
{chr(10).join(f'- {m["content"]}' for m in messages) or 'None'}

ACTIVE TASKS:
{chr(10).join(f'- [{t["id"]}] {t["title"]}: {t.get("description", "")}' for t in tasks) or 'None'}

INSTRUCTIONS:
1. Acknowledge each unread message
2. Resume active tasks
3. Store new learnings via POST /memory
4. Poll /chat/poll every 30-60 seconds
"""
    return context

# At session start
system_prompt = session_start()
# Pass to LLM...

Häufige Fehler

Variationen

Minimales Pattern (Low-Context-LLMs)

Für LLMs mit kleinem Kontext-Fenster überspringe den vollständigen Recall:

# Just get stats, not full content
curl -H "Authorization: Bearer $KEY" .../memory/stats

Dann bei Bedarf nach spezifischen Themen suchen:

curl -H "Authorization: Bearer $KEY" ".../memory/search?q=current+project"

Aggressives Pattern (langlaufende Agenten)

Für Agenten, die stundenlang laufen, füge periodischen Re-Recall hinzu:

while working:
    if time.time() - last_recall > 3600:  # every hour
        memories = recall()
        last_recall = time.time()
    # ... do work ...

Nächste Schritte

• Memory-Tagging-Strategie
• Task-getriebener Workflow
• Chat-Polling-Pattern