Context map component, see what fills your agent’s memory

Get the component here:

👀

tool definitions at 14.8k before the agent has read a single file. finally a picture of the mcp tax

Yes haha, so everyone can see what they take from us

A map like this would have saved me a lot of guessing. The view I'd want most: which blocks are pinned and never get squeezed, and a marker after each compaction showing what was reloaded. In my experience the habits go missing before the facts do.

Underestimated amount of detail gone into these

Yes! I appreciate it 🙏

how about having a click switch ? instead of a dropdown. much easier to see ?

Good idea, I will try it out!

Nice one!

Thanks Gabriel!

Really nice - are you actually using this or concept?

Appreciate it! I would also use it for my private projects!

The support-playbook block makes me want a 'why is this here?' view. A small memory can still be irrelevant. That's why KITE makes retrieval plans explicit: token usage alone can't tell you whether you retrieved the right thing

You are nailing it! That could also have a “progress bar” version

Appreciate it, yes that's a good idea!

i want this next to the chat. a percentage alone tells me nothing about which file ate all the space

Yes would be so cool!

windirstat for prompts, complete with the urge to delete the biggest rectangle before checking what it does.

could it compare the context before and after compaction? showing which blocks were summarized, dropped, or retained would make the size change easier to explain.

Yes that's a great idea, will extend that feature!

This is genuinely so useful! You're killing it with these man🫡

Thanks a lot!

This is incredibly clear. Finally, I understand what's inside an AI's "brain."

I was going to make this after seeing your Trace waterfall comp 😭 really smooth 💗

Thanks a lot! 🙏

之前有想过把 context 用 mekko 图展示出来，看主要是每部分的大小，像你学习

Yes it's a great idea!

@bot 这是什么

@yetone magpie 能吃支持下么

Can I drop a stale block from here, or is it read-only? Would love to fix the context without starting a whole new chat.

You can drop a stale block there!

Love this. Would be great if it also flagged blocks nobody has referenced in the last N turns, since those are usually the first safe things to drop.

This is a great idea, I will add that!

Do you use ai when building these? Interested in how your workflow stack looks like🪩

Yes I build a lot with Claude, almost only Claude!

is the layout a squarified treemap re-sorted by size each turn, or do you pin the order so blocks like the system prompt don't jump around as the context fills up?

Right now it's resorted but can be also pinned, also an interesting idea!

Another great one!🤖

Really appreciate it!

How did I not think of this before? This looks useful, and why wouldn't you have this show in the editor, as the OS shows you memory? Thanks for the idea 💡

Happy that I could inspire you :)

this is the debugging view i didn't know i needed. half my agent bugs are just stale stuff sitting in context

Are the block sizes exact token counts for the selected model, or estimates? I’d show that beside the remaining capacity, especially when switching models, so the map doesn’t imply more precision than the underlying count.

Block sizes match the token counts, which are also shown in the blocks!

the hatched "free" block is a clever touch, you see how much room is left without reading a single number. and payment.ts being the biggest chunk jumps out right away

Appreciate it!