Cursor just one shot OpenAI Codex and Claude Code with their new update. they solved Context Decay.... ok vibe coder I know you don't know what that means, but just trust watch.

codex watching this video

Curious what "solved" actually means here technically, since what I can see from Cursor's own changelog is Projects, launched just yesterday, which delegates to subagents and maintains context over months, that's a real feature, but maintaining context and eliminating context rot aren't the same claim, and I'd want to see how it actually holds up at scale before calling it solved. This is something I've been researching and living with hands on for close to two years now. Context engineering is the actual skill underneath all AI coding work, without it you don't get what you planned for, you get whatever the model happened to retain. The mechanism is specific: once you cross roughly 80% of a context window, around 800k tokens in a 1M window, the model starts losing grip on what happened at the start of the session. That's the actual moment context rot sets in, and it's when you need to migrate to a fresh session. Doing that manually means writing a handoff file plus foundational documentation for the next agent, and it's genuinely an art form, too little and the new session is lost, too much and you've burned a chunk of the fresh context window before real work even starts. That exact problem, getting the amount and shape of carried-over context right across sessions, harnesses, and models, is what I've been building The Curator to solve directly. Nice presentation of what's possible either way, video style's a bit much for my taste but the demo itself is compelling.

claude code 和codex不也有记忆么？

Context decay is the real bottleneck

How does Cursor actually handle context decay in this update? Curious how it holds up against Codex and Claude Code in practice.

Context decay is what kills my overnight agent runs—not raw model IQ. If Cursor actually fixed that, the Claude Code comparison finally has a real axis.

Love Projects they feel way different to everything else Kind of seems like what I thought graph engineering promised to be But not quite sure why this same thing is not possible with a graph underneath Can agents not just store / manage context like this with a graph or is there some reason why Projects gets over hurdles there as well?

Context decay is not solved by one impressive pass—the real test is whether the agent can preserve decisions across a messy repo and the next session.

the part that hit me wasn't the tokens disappearing, it was the model losing WHY those early decisions were made. summary re-injection at chunk boundaries doesn't recover that.

I don’t mind the edits but not a fan of the sfx

Hype much?

Context decay is just your agent forgetting the brief mid-PR. We gave it a name so we feel less insane.

🤌

So good!

Love the editing… entertaining 😂

Context decay is the failure mode people notice only after a session has already gone off course. The real test is whether the tool can keep the current files, decisions, and last verified result visible enough that the next step starts from reality instead of stale assumptions. That is a much more useful upgrade than simply claiming a larger context window.

If Grokbot isn't taking away manual context limit management, then this is just another agent

Context decay is the silent killer of long agent sessions. If your agent forgets the architecture halfway through, no model upgrade saves you. Persistent workspace beats a bigger context window.