Memory and Context

Agents have a 3-tier memory system: working memory (current), summaries (history), and archives (long-term).

​

Memory Tiers

​

Working Memory

Current conversation buffer. Keeps recent messages in context.

• Default: 8,000 tokens
• Contains: Last N messages
• Cost: Included in each request

​

Summary Memory

When working memory fills, creates summary of key points.

• Default: 5 messages
• Triggered when: 5+ messages in working
• Contains: Condensed history

​

Archive

Older summaries stored separately.

• Default: Keep all
• Accessed when: Needed for context
• Cost: Additional retrieval tokens

​

Configuration

​

Adjust Working Memory

Smaller (4,000 tokens):

• Pros: Lower cost, faster responses
• Cons: Forgets details quickly

Larger (16,000 tokens):

• Pros: Better context, fewer summaries
• Cons: Higher cost, slower

Default (8,000 tokens):

• Good balance for most uses

Change:

1. Agents > Select agent
2. Settings > Memory
3. Set working_memory_size
4. Click Save

​

Summary Threshold

When working memory reaches N messages, create summary.

• Lower (3): Summarize more often (cheaper)
• Higher (10): Summarize less often (better context)
• Default (5): Good balance

Change:

1. Settings > Memory
2. Set summary_message_threshold
3. Click Save

​

Memory Management

​

View Memory Usage

1. Agents > Select agent
2. Analytics tab
3. See memory tokens used

​

Clear Memory

1. Open conversation
2. Click Clear History
3. Confirm

Conversation history deleted. Cost savings.

​

Export Memory

1. Open conversation
2. Click Export
3. Download as JSON/PDF

​

Best Practices

​

Balance Cost and Quality

​

Monitor Memory Growth

Check usage dashboard:

• How much memory used per conversation?
• Trending up or stable?
• Adjust if needed

​

Archive Old Conversations

Periodically clear old conversations to reduce costs.

​

Memory Limitations

• Working memory is conversation-scoped (not shared between agents)
• Summaries are automatic (can’t manually create)
• Archives are read-only (can’t modify)
• Memory persists even after agent updates

​

Troubleshooting

​

Agent forgets context

• Increase working_memory_size
• Increase summary_threshold
• Ask user to provide context again

​

Memory costs too high

• Reduce working_memory_size
• Lower summary_threshold (summarize more)
• Clear old conversations

​

Incorrect summaries

• Summaries are LLM-generated (won’t be perfect)
• Verify important details in conversation
• Consider manually archiving if critical

​

Next Steps

• Creating Agents
• Model Routing
• Usage & Limits