It remembers nothing
Every call starts from zero. Memory is your job.
The API keeps no memory between calls. If you want the model to know something, you send it, every single time.
Call 1
user
Hi! I'm Priya, final year at COEP, preparing for Amazon.
model
Hey Priya! Great to meet you, and a massive shoutout to COEP—such a legacy institution with an incredible tech pedigree. Cracking Amazon i…
Call 2: what you send
user
What's my name, and which company am I preparing for?
The reply
I don't have access to your personal information, past chat history, or context about who you are, so I don't know your name or which company you are preparing for! If…
Real data from gemini-3.5-flash-lite, 2026-10-16
So an agent resends the whole conversation on every call. Watch the context jar fill up, call after call.
- model calls
- 1
- tool calls
- 0
- tokens in
- 170
- tokens out
- 21
- time
- 1.9s
Task My order A123 hasn't arrived yet. Where is it?
Model call 1. It read 170 tokens and wrote 21. It asked for a tool, so it isn't done.
get_order({"order_id":"A123"})
stop_reason: tool_use
Context window
170 tokens sent
- Your messages12
- Tool definitions124
- System prompt34
0.02% of the 10,48,576-token window. Every call sends all of it again.
Context window170 tokens sent
- System prompt34
- Tool definitions124
- Your messages12
Why does an agent's 10th model call cost more than its 1st?