Antigravity compaction dropping details? Save a handoff file first, then run /compact

Antigravity compaction dropping details? Save a handoff file first, then run /compact

If an Antigravity agent forgets something you settled an hour ago during a long task, compaction is one place to look. Google's docs say Antigravity automatically compacts older turns into checkpoint summaries. That is the documented mechanism. What gets lost is a separate question, and the evidence there comes from users: forum posts describe summaries that recast a solved problem as open, or dropped constraints like "don't touch this file."

What follows is based on the docs and changelog as of November 13. Two things are in your hands: when you trigger a compaction yourself, and where your task state lives outside the chat.

In the CLI, check /context first

The Antigravity CLI ships a /context command to inspect the active conversation's token consumption and context window usage. It breaks usage down by user messages, agent responses, tool calls and more, and lists a Checkpoint buffer row: the tokens held by summaries standing in for older, already-compacted turns. If the agent drifts after a compaction, the docs point you to /rewind to roll back to an earlier checkpoint.

One catch: that page lists the Antigravity CLI as the only place the command is available.

In the app, you can compact, but the meter is unconfirmed

The desktop app added /compact in v2.24.0 on October 15, 2026. The changelog describes it as summarizing a long conversation to free up room for the agent. The upside is timing: you can compact right after a phase wraps up and the state is clean. Triggering it by hand does not guarantee a more complete summary, though.

Here's the annoying part. The only usage-related entry in the app changelog is v2.18.1, which changed the existing token usage breakdown to show rules and other customizations against their own budgets; rules get a 20,000-token budget. That entry covers customizations, not the conversation as a whole, and nothing through v2.32.0 on November 13 confirms a conversation-level usage display in the app. Undocumented isn't the same as absent. If you can't find such a reading in the app either, you can make a rough call on when to compact or start fresh from turn count, tool-call volume and where the task stands, but none of that substitutes for an actual token usage reading.

Save state first, then compact

On the Google Developer forums, an embedded engineer described a summary that recast a solved problem as still open, so the agent didn't realize the code on disk was the hard-won fix. His answer is a local Markdown journal that the agent treats as the source of truth instead of its conversational memory, and he treats every compaction as an invalidation event.

The order matters. While the agent's memory is still intact, keep a handoff file in the repo listing verified fixes, files that must not change, and next steps, and check it yourself. Add a rule telling the agent to read it before editing code. Only then run /compact. If you ask the agent to write the file after a compaction, you may be saving a summary that has already drifted.

After compacting, have the agent restate the file and compare its version against the original before it touches any code.

Not everyone likes this. The author of another forum thread said he wanted a built-in /compact command, not another Markdown file in the repo. The app has delivered the command. The changelog doesn't say whether /compact includes any way to verify the summary, so our advice is to keep the handoff file anyway and check the agent against it yourself after each compaction.

Share this article