OpenAI Codex Chronicle memory feature launch

Chronicle is philosophically close to Microsoft Recall: a background agent that captures your screen, OCRs it, and stores summaries so an assistant can reason over what you've been doing. But the implementation diverges in a consequential way. Recall invested heavily in an encrypted, hardware-bound local database after its rocky debut; Chronicle writes memories as plain-text Markdown under ~/.codex/memories_extensions/chronicle/, readable by any local process that can see the user's home directory.

Short-lived screen capture files live in $TMPDIR/chronicle/screen_recording/ and are garbage-collected after 6 hours, but the distilled memories persist. For a product targeted at developers — who routinely have API keys, private repo paths, and production incident details on screen — unencrypted at-rest storage is a notable choice. It is also why the feature is gated to $100/month Pro users on Apple Silicon with macOS 14+: OpenAI is limiting the blast radius while it iterates.

Analyst Aakash Gupta's read is that Chronicle exists because the memory layer has overtaken raw model quality as the differentiator between coding agents. If every major lab can ship a frontier model, the agent that remembers what you were doing yesterday — which branch, which stack trace, which Figma frame — wins the next session.

That framing explains an otherwise odd economic choice: OpenAI openly warns that Chronicle 'burns through rate limits quickly' because sandboxed Codex sessions continuously summarize captures in the background. OpenAI is effectively subsidizing always-on summarization compute to build a persistent context graph of each Pro user's work. Against Anthropic's Claude Code — the competitive foil cited in TechCrunch's framing of OpenAI's April 16 Codex expansion — that persistent graph is the asset model swaps cannot replicate.

Chronicle asks developers to grant Screen Recording and Accessibility permissions and let a background agent watch their monitor in exchange for escaping the chore of restating context. OpenAI is transparent about the trade: the docs explicitly flag prompt-injection risk because any rendered web page or chat window Codex sees becomes eligible input, and recommend pausing Chronicle before meetings or when viewing sensitive material.

The feature also uploads screens to OpenAI servers for OCR and summarization — OpenAI commits to not retaining them post-processing or training on them, but the pipeline still constitutes continuous exfiltration of whatever is on screen. Developer YouTube explainers note the pause/resume control lives in the macOS menu bar, which makes the opt-out ergonomics acceptable but puts the burden of discretion squarely on the user.

Chronicle's unavailability in the EU, UK, and Switzerland is not incidental. The Next Web reads it as tacit acknowledgment that continuous screen capture, even with on-device memory persistence, sits awkwardly against GDPR purpose-limitation and data-minimization norms. That regional split is becoming a familiar pattern for ambient-memory products.

It also feeds a counter-current visible in community reaction: developers are surfacing open-source analogs that perform similar screen-to-memory capture and arguing that a proprietary vendor launch doesn't foreclose the category — it legitimizes it. The contrarian read is that auditable, self-hostable screen-memory tools become more attractive precisely because Chronicle is closed, server-dependent, and geographically gated. OpenAI has validated the problem; whether it owns the solution for privacy-sensitive and non-US developers is an open question.