ChatGPT's Memory Gets Smarter, and Figma Charges for AI

OpenAI — ChatGPT "Dreaming" Memory System#

Source: https://openai.com/index/chatgpt-memory-dreaming Credibility: High (first-party product announcement)

What happened: OpenAI shipped a new memory architecture for ChatGPT called "Dreaming." The system is designed to keep memory fresh and relevant across conversations rather than just accumulating facts indefinitely. The name implies an offline consolidation process — preferences, patterns, and context get synthesized periodically rather than stored as raw recall. This is a user-facing product change for ChatGPT, not just an API update.

Key capabilities:

• Memory consolidation to keep context fresh and relevant (not just additive)
• Synthesizes preferences and patterns across past conversations
• Framed as making ChatGPT "more helpful" by retaining what matters and deprioritizing what doesn't
• Available in ChatGPT (tier details not specified in the excerpt)

Why it matters for PMs: Memory architecture is one of the most important unsolved UX problems in AI products. Most memory implementations today are essentially append-only logs — they get noisy fast. If "Dreaming" actually consolidates and prioritizes context the way the name implies, it's a meaningful signal about how the category is thinking about long-term user relationships. For PMs building products on top of ChatGPT or thinking about memory in their own AI features, the design question shifts from "what do we store?" to "what do we forget, and when?"

Critical questions:

• What's the consolidation mechanism — is it model-generated summaries, user-visible memory entries, or something opaque? Transparency here matters a lot for user trust.
• How does this interact with user control? Can users see what got "dreamed away"?
• Is this rolling out to all tiers or just Plus/Pro? The segment matters for adoption signal.
• Does this change the behavior of the API memory tools, or is it ChatGPT-only?

Action you could take today: If your team is building any feature that involves user preferences or history, sketch out the "what do we forget?" side of the memory design — not just what you store. OpenAI just made that the competitive question.

Figma — Pay-as-You-Go AI Credits and "Check Designs" Launch#

Source: https://help.figma.com/hc/en-us/articles/35865276858647-Manage-AI-credits (pricing) and https://help.figma.com/hc/en-us/articles/39592284074263-Check-designs-in-Figma (feature) Credibility: High (first-party changelog entries, June 3-4, 2026)

What happened: Figma shipped two things back-to-back this week. First, Professional plan admins can now purchase additional AI credits on a pay-as-you-go basis — breaking AI usage out of flat subscription caps for the first time at that tier. Second, they launched "Check designs," a feature that compares your designs against your design system, flags what's off, and suggests fixes. It's positioned as a pre-handoff quality check.

Key capabilities (Check designs):

• Compares active designs against the connected design system
• Flags inconsistencies automatically
• Surfaces suggested corrections
• Targeted at design-to-dev handoff quality

Key capabilities (Pay-as-you-go AI credits):

• Available on Professional plan (previously only Enterprise had flexibility)
• Admins purchase credits in addition to base plan allocation
• Enables teams to scale AI usage without upgrading the entire plan tier

Why it matters for PMs: These two changes together tell a clear story. Figma is betting that AI credit consumption will grow fast enough to warrant a new monetization layer, and they're lowering the barrier to access it before teams hit limits and churn. The "Check designs" feature is also interesting as a quality enforcement tool — it's the kind of thing that could meaningfully reduce PM-to-designer review cycles if it catches style drift automatically. For PMs who run design reviews, this is worth testing: it surfaces issues before you get to the meeting.

Critical questions:

• What's the credit pricing? Without that, it's hard to evaluate whether the pay-as-you-go model is favorable or punitive.
• How accurate is "Check designs" against complex or evolving design systems? If it flags too many false positives, teams will ignore it.
• Does this apply to Figma Make (AI prototyping) usage, or only to traditional AI features like Rename Layers, AI search, etc.?
• Is this a step toward usage-based pricing across the whole product, or just an AI feature exception?

Action you could take today: If your team uses Figma on the Professional plan, have your admin check AI credit usage for the past 30 days. With pay-as-you-go now available, you can either budget more accurately or use this as a signal for whether upgrading to Org/Enterprise makes sense.

Cassie Kozyrkov — CEOs Are Betting on AI-Driven Layoffs. The Data Isn't.#

Source: https://medium.com/swlh/ceos-are-betting-on-ai-driven-layoffs-the-data-is-betting-against-them-5b8c11ad056f Credibility: Medium-High (Cassie Kozyrkov is former Chief Decision Scientist at Google; this is a data-driven opinion piece, not a study)

What happened: Cassie Kozyrkov published a piece arguing that the CEO narrative around AI-driven workforce reduction is running ahead of the evidence. The core argument: while executives are publicly committing to headcount reductions tied to AI adoption, the actual data on productivity, employment, and task substitution tells a more complicated story. Kozyrkov frames this as a decision-making failure — executives making irreversible bets on reversible assumptions.

Key patterns:

• AI tools tend to augment workers rather than replace them in most documented cases
• The "productivity gains = fewer people needed" logic contains a hidden assumption about demand elasticity that often doesn't hold
• CEOs are conflating "AI can do this task" with "AI will replace the person doing this task" — those are different claims
• The data on actual AI-driven employment reduction lags the rhetoric significantly

Why it matters for PMs: This is directly relevant to how you frame AI features to executives and how you build them. If your org is making staffing decisions based on what AI might eventually automate, you're building roadmaps on assumptions that don't have strong empirical backing yet. More practically: if you're shipping AI features for internal teams or customers, the "this will reduce headcount" framing is politically attractive but often wrong — and building around it can lead to features that optimize for the wrong outcome.

Critical questions:

• Kozyrkov is making a normative claim (CEOs are making bad bets), but what's the time horizon? Over 10 years, the productivity displacement story may be different.
• Which sectors or task types are the exception? "AI isn't replacing workers" may be true in aggregate while being false in specific domains.
• Is the lag between rhetoric and data a signal that displacement is coming but delayed, or that it won't happen at the predicted scale?

Action you could take today: If your team is building internal AI tools, check how success is being defined. If "reduced headcount" is on the KPI list, that's worth a direct conversation about whether the evidence supports it — or whether you're being set up to ship something that will be called a failure regardless of actual productivity impact.