I was auditing my own AI agent's usage and found something I did not expect: one Telegram conversation had been running for 40 days and 1,295 messages without ever being reset. It felt like a single continuous chat — until I looked at the numbers. That long-lived session was quietly costing me more on every single turn, and the quality at the end was worse than at the start.
This is not an edge case. It is the default behavior of every chat-with-an-agent unless you consciously break it. Here is how it works, why it matters, and how often you should hit reset.
What a "session" actually is
Every time you talk to an agent, the whole conversation so far is sent with your message so the model remembers the context. In agent terms this is the context window, and a session is the span of messages that keeps growing until someone starts a new one.
In my case the session started on July 17 and only ended on August 26. Every single message over those 40 days carried the entire history with it.
The two costs of a long session
It gets more expensive. The model bills on input tokens — and every turn re-sends everything you have ever said in that session. A 5-message conversation sends 5 messages of context. A 1,295-message conversation sends 1,295 messages of context, every time. Even with prompt caching (which makes repeat reads cheaper), the base cost keeps climbing as the session grows. By the end mine was pushing 50,000+ input tokens per message — just to say hello.
It degrades in quality. Context is not free attention. As a session fills with old, unrelated back-and-forth, the model has to wade through it to find what matters. Stale assumptions from weeks ago start to steer the answers. You notice it as the agent "forgetting" the real goal, repeating itself, or picking up a topic that you dropped long ago. That is usually not the model misbehaving — it is a bloated context steering the whole conversation.
When to start a new session
There is no fixed timer. The rule is: one session per task, not per week.
- You change task or topic — reset. This is the single biggest win.
- You passed a lot of files or a long spec that is no longer relevant — reset.
- Input tokens are creeping up (roughly past 50,000 is a good cue) — reset.
- The agent seems off — repeating itself, missing the point, stuck on old context — reset and restate the goal cleanly.
- You want to drop a model or provider override — reset.
For most people this lands at a handful of resets per day, not one per month.
How to reset
In Hermes the command is /new (alias /reset):
/newThat ends the current session and starts a fresh one. You can also give it a name:
/new my-task-nameA clean session is cheaper per turn and gives the model a clear slate. Your persistent memory and skills survive the reset — those are stored separately and re-injected on every fresh session — so you do not lose durable facts. You only drop the accumulated conversation baggage.
The takeaway
A long-running chat feels natural — it is one continuous conversation to you. But underneath, the agent is re-reading months of history on every message. My 40-day session was a 1,295-message mistake in waiting. Starting a new session when you move to a new task is the cheapest, simplest thing you can do to keep an agent both fast and sharp.
Filed under the practical side of using AI agents — run /new often, and your agent will thank you.
