feature
Oct 31, 2026Claude Code, Codex CLI, Cursor Agent, and Antigravity now share local context about your own past sessions — zero added latency, built on data no code-search tool has.
It's a new opt-in mechanism, and it's the first thing in the product built on data nothing else has: not "where is the code" — repo maps and symbol indexes already do that — but "which of my own past sessions touched this file, for what reason, alongside what else."
At session start: a tiny layout card — top-level directories, sibling repos, the newest migration file if your project has one — built from nothing but your local filesystem. No network call, no judgment calls, just facts about what's on disk right now.
On a prompt that looks related to something you've worked on before: a short "similar past work" note — which prior sessions probably relate, which files they touched, which files tend to change together with those, and whether any of them have changed since. This is scored entirely against a small local cache on your machine. The only network traffic involved is a periodic background sync — never on the path of your actual prompt, so there's no latency hit at the moment you're waiting on a response.
Both only show up when there's a real signal. A weak or empty match means nothing gets injected — the point is to remove a discovery step, not add text you have to read past.
The interesting engineering problem here was making this work the same way across four tools that don't share a hook model. Most can inject this note the instant you submit a prompt. Cursor's prompt-submission hook can't inject anything at all, so there it's captured at that moment and delivered at the very next opportunity in the same turn — same result, one hook later. One shared core does the actual ranking and caching for every tool; each one just gets a thin adapter for how its hooks work.
Visibility follows the same rules search already uses across your team or org — scoped per person per project, server-side, at the widest level you're already allowed to see. Turn your own session sharing off, and your sessions simply don't surface for teammates here either, same as everywhere else in the product.
One opt-in flag on the existing hooks installer. If you already have ReclaimLLM's hooks installed, re-running the installer with that flag on turns this on without touching anything else you've configured.
Every coding-agent session on a nontrivial task starts the same way: the agent doesn't know your repo yet, so it spends several turns on ls, grep, find, and reading files just to orient itself — then re-sends all of that exploration as context on every later turn for the rest of the session. In one session we measured, 70 of 188 total tool calls went to pure discovery before a plan was even ready, and that discovery context accounted for over 15 million cache-read tokens over the session's lifetime, purely from being re-sent turn after turn. None of that is new work — it's the agent re-deriving something a past session on the same project had already established.
We already capture the thing that removes that step: which sessions touched which files, for what intent, together with what else. Start Brief is that history, surfaced before the agent has to go looking for it — and because it's local and hook-injected rather than a tool the agent has to decide to call, it costs nothing when it doesn't fire.
This is new, so it's off by default and deliberately conservative — we'd rather it say nothing than say something low-confidence, so expect it to stay quiet more often than you might like at first. We also don't show a savings number for it anywhere yet; we don't have an honest way to measure what would've happened in a session without it, and we're not publishing a number we can't stand behind. That's also the main thing that would change its default.
A few other things worth knowing: the first time you use this in a project, project-name matching may still be catching up, so it can stay quiet until it's synced once. Cursor and Antigravity both have documented, historical reliability issues with the specific hook channel this depends on, separate from anything we control — we check for delivery as best we can, but that check can only suggest evidence, never prove it. And the admin-side status view for this currently only reports Claude Code; the other three tools don't show up there yet.
If you turn this on and it's too chatty, too quiet, or surfacing the wrong thing, tell us — that's exactly the feedback that decides whether this becomes a default instead of an opt-in.