OpenAI is expanding Codex’s “memories” experiment with a new feature called Chronicle, aiming to reduce how often context needs to be repeated in day-to-day coding work, according to a post from OpenAI Developers.
OpenAI said it released a preview of memories in Codex last week, and Chronicle now “improves memories using recent screen context,” so Codex can help with ongoing work “without” restating what’s been happening.
What Chronicle adds to Codex memories
OpenAI’s description focuses on reference resolution—helping Codex understand vague pointers like “this” or “that” by using what’s currently visible on screen. The examples OpenAI gave include:
- an error that’s on screen
- a document that’s open
- “that thing” from a couple weeks earlier
Over time, OpenAI said Chronicle helps Codex learn patterns such as the tools being used, projects being revisited, and workflows being relied on.
How it works, and what OpenAI says is stored
OpenAI said Chronicle “runs background agents to build memories from screen captures,” and that it “uses rate limits quickly.”
In terms of storage, OpenAI described a local pipeline: screen captures are “stored temporarily on device” to generate memories, and those memories are also “stored on device.” OpenAI added that memories can be inspected and edited, while cautioning that “other apps may access these files.”
Availability: Pro on macOS, with regional exclusions
OpenAI said the rollout is starting with Pro users on macOS, excluding the EU, UK, and Switzerland, while the company learns where the feature helps most and iterates on the experience.
Early reactions: utility vs. surveillance and security concerns
The announcement quickly drew comparisons to Microsoft’s Recall, with one developer calling it “similar to MSFT’s Recall.” Others focused on practical benefits: multiple replies highlighted how much friction comes from repeatedly re-onboarding an assistant with the same working context every session.
At the same time, concerns clustered around monitoring, privacy, and safety. One reply asked whether it is “always watching,” another called the approach “risky” and argued for screen sharing only when explicitly enabled, and another raised the potential for a larger “blast radius” if an agent with continuous screen access is exposed to prompt injection vulnerabilities.
A more pragmatic complaint also showed up early: one Pro user said Chronicle “really chewed through” plan limits, characterizing the feature as “research preview” until usage is smoothed out.
