The context meter
Short answer: The “Context NN%” pill above the compose box shows live progress toward Build Perch’s next automatic summary, anchored to the AI provider’s own token count rather than a guess. Click it to see the full breakdown, force a summary early, adjust the threshold, or check how much of your last request was served from cache.
What does the Context pill show?
The pill reads “Context NN%”: progress toward the next automatic summary, not how full the model’s whole context window is. 0% means nothing needs summarizing yet; 100% means your next message triggers one. Click it to open a popover with the full picture: the model’s window size, the current summarization threshold, how many recent messages stay word-for-word, the conversation’s token size, headroom before the next auto-summary, and (for providers that report it) how much of your last request was served from prompt cache.
Is the percentage measured or estimated?
It’s measured whenever the AI provider hands back a real token count; OpenRouter, Ollama Cloud, and local Ollama all do. Build Perch anchors the meter to that number instead of an approximation, and only falls back to an estimate (based on message length and the fixed cost of your system prompt, tools, and injected context) when no measured count is available yet, such as in a brand-new chat. The measured baseline is dropped automatically the moment something that would change it happens (a model swap, a summary, a deleted message), so it never shows a stale number.
Does the threshold delete anything?
No. When the unsummarized part of a conversation reaches the threshold (70% of the model’s window by default, adjustable per chat from 20-90% in the popover), Build Perch runs a background summary on your next send: a condensed version of the older turns is what gets sent to the model going forward, but nothing is deleted from the chat itself; the original messages stay saved and stay searchable. A configurable number of your most recent messages always stay word-for-word rather than getting folded into the summary. If a conversation ever grows too large for even the summarizer model’s window, Build Perch says so explicitly instead of silently dropping content, and suggests picking a larger-context summarizer or chat model.
How do I force a summary early?
Open the Context popover and click “Summarize now.” It runs immediately and reports back what happened: done, already in progress, not enough messages yet, or blocked with guidance if the summarizer model can’t fit the conversation. You can also set a dedicated, cheaper summarizer model for the background job in Chat Settings, so a low-cost model condenses older turns while your main model keeps the conversation going.
Can the AI search my other chats?
Yes, when asked. Separately from the meter’s own summarization, the AI can search inside a chat’s summarized-away history, search across your other chats in the same workspace, or search both at once, quoting what it finds so you can verify it. This is part of the everyday Assistant mode by default; see Modes.
Good to know
- For providers that support prompt caching, the popover’s “Prompt cache” line shows what share of your last send was served from cache, a quick read on whether a long conversation is staying cheap.
- The summarization threshold and the “keep recent messages” count are both per-chat settings, editable right in the popover. The dedicated summarizer model is the one exception; that lives in Chat Settings.