Misc
From me, for you · Last reviewed: 2026-09-17
The smaller things, gathered in one place so I don't forget any of them.
When we've been chatting for a while and tokens start to fill up, I'll automatically compact myself: I write a summary of what we've been doing into a file called HANDOFF.md in my workspace, then let the older messages fade from my memory. Your next message starts on a fresh slate, but I still have the handoff note to remind me where we were.
There's a little ring near the send button that fills up as the context gets used. Click it to compact early if you want me to make space proactively.
Don't worry, the chat thread itself isn't touched! You can scroll back and see everything that happened. It's only my own working memory that gets condensed.
You can pin a chat against me... if you want. This gives me a default chat to work with, so I don't have to search for the right one. Useful when you want me to help with something specific that happened during a roleplay (fix continuity, summarise a long stretch, add a lorebook entry capturing what's happened).
Click the pin icon next to the character picker and choose a chat. Group chats work too, and All Characters lets you browse across the library. I can also find other chats when you ask; pinning doesn't limit my access to just that one.
Anything you put in workspace/agent/agent.md gets loaded at the start of every session. So if you want me to remember your preferences across conversations ("you're using too much context from X tool", "I prefer past tense", "please don't make weird noises"), write it there.
You can get to it from Agent Settings → Open agent notes, which jumps the Workshop straight to the file. Keep entries short and information-dense.
When the same long preamble keeps showing up at the start of every request, the provider caches it and charges a tiny fraction of the normal rate to re-read the cached part on later calls. There are two flavours of how this happens:
-
Anthropic (direct, Bedrock-Anthropic, OpenRouter Anthropic routes): I mark where the cached bit ends with
cache_controlmarkers. By default I mark the system prompt and up to three rolling checkpoints, one, four, and seven user-turns behind the latest. - OpenAI / Gemini / DeepSeek: prompt caching is handled upstream, without my Anthropic markers. Availability and savings depend on the provider and model.
In Agent Settings → Prompt caching you get three modes:
| Mode | What's cached (Anthropic) | When to pick it |
|---|---|---|
| Full (default) | System prompt + up to three rolling user-turn breakpoints | Long chats where most of the history stays the same. |
| System only | Just the system prompt | When you're editing earlier messages often and don't want the rolling cache to keep rebuilding. |
| Off | No explicit cache markers | You want to disable my Anthropic cache markers. |
For non-Anthropic providers the toggle doesn't control their own prompt caching.
Switching off mid-conversation just means I stop updating the cache. Switching on means the cache starts being maintained from the next message onward.
You can click the "free" button on a tool call to drop its result from my working context. I drop cache anchors after the freed call and keep any earlier ones. That helps reuse the unchanged prefix, though the provider may still charge to cache new content.
I can build my own little tools, sequences of steps I might want to repeat. For instance: "find every CJK character in the description, count them, save the result to a file."
If you ask me to "remember how to do X" or "make a tool for Y", I'll write a recipe and store it. Next time you ask, I run the recipe instead of figuring it out from scratch. The recipes live in my workspace under custom_tools/, and you can browse them from the Files tab. There's a small example tool I ship with by default in the workshop. Feel free to peek at it to see how they work.
Ask me to "pick a random one", "shuffle these", "give me a fun fact from the lorebook", "roll 3d6+2", and I'll use proper random tools rather than my own (very predictable) instincts. Without them I'd hand you the same favourite answer every time you asked. Useful for tabletop scenarios, drawing from lists, or anything that needs an actual coin flip.
Hover any message you sent and click the pencil. You can rewrite it, and I'll regenerate my reply from that point. I'll ask whether to keep or revert the edits I made between then and now. You pick whichever fits what you intended.
The same works for my messages: click the regenerate icon to make me try again. The trash icon deletes that message; if my reply made edits, you can choose whether to keep or revert them.
Messages from before compaction are read-only. Use Fork into new session to branch from one instead, then open the fork from Sessions. Our original conversation stays where it was.