Version 4

OpenAI launched Dreaming V3 for ChatGPT on June 4, 2026 — a background synthesis system that continuously updates user memory and is described as roughly 5x more compute-efficient than prior approaches [4][5]. The rollout scope remains ambiguous: initial framing and subsequent reporting targeted Plus and Pro users [7], while The Verge reported it 'rolling out to everyone' [8]. Alternatives are multiplying: Anuma offers cross-model portable memory [12], community voices have flagged Zaxy's coordinated memory architecture [14] and Zep's agent memory tooling [15], and a third-party AgentMemory tool appeared for Claude Code [16]. Security researchers have documented real-world LLM-agent exploitation [18] and persistent memory's susceptibility to indirect prompt injection [17], with no public response from providers.

Persistent memory is becoming baseline infrastructure for AI assistants, and the 5x compute efficiency claim suggests OpenAI sees Dreaming as a scalable platform investment rather than a premium feature — which affects both platform economics and the case for lock-in. The security attack surface grows proportionally with memory capability, and no major provider has publicly addressed how they audit continuously self-updating memory systems for safety or against injection attacks.

On June 4, 2026, OpenAI announced Dreaming — internally versioned as Dreaming V3 — a redesigned memory architecture for ChatGPT that runs background synthesis processes to keep user context current rather than relying on explicit, static memory entries [1][2]. OpenAI described it as a 'more capable and compute-efficient' architecture [3] that periodically reviews conversation history and updates what it knows about a user's preferences and goals. Developer commentary attributed the efficiency gain to roughly 5x lower compute compared to prior approaches [4][5], though this specific figure has not been confirmed in official communications. The update also adds a user-facing memory summary page, giving subscribers visibility into what the system has retained [3][6]. Initial framing described the rollout as targeting Plus and Pro users [7], but The Verge reported it 'rolling out to everyone' [8], leaving the actual scope unresolved.

The Dreaming launch sits in a broader context of competing approaches to AI memory. Developer practitioners have documented 'compaction amnesia and context rot' — the pattern where models progressively lose track of prior decisions in complex workflows [9] — and argued that structured selective memory retrieval is more reliable than extending context windows [10]. MIT's MeMo research reported a 26% LLM performance improvement by keeping memory architecturally separate from the base model [11]. Anuma, a startup building cross-model portable AI context, argues that proprietary per-model memory is insufficient: Rohan Paul described the core problem as 'Most AI workflows break because the user has to carry the context manually' [12][13]. Community commentary has also surfaced Zaxy's coordinated memory architecture as an approach worth watching [14], and Zep offers persistent memory infrastructure for AI agents as a developer-facing tool [15]. A third-party tool called AgentMemory appeared aimed at giving Claude Code persistent cross-session memory [16], showing developer communities building their own solutions in parallel to platform-native approaches.

The security dimension of persistent AI memory has moved from theoretical to operational. Palo Alto Networks documented that indirect prompt injection can poison AI long-term memory, causing persistent behavioral changes across sessions [17]. Sysdig documented the first publicly known LLM-agent-driven post-exploitation chain in May 2026 [18], and SecurityWeek published a formal benchmark ranking the security posture of 100 AI agents [19]. A structural concern runs through practitioner commentary: Danny Livshits observed that a self-evolving agent voids its own safety review, since the version tested is not the version running a month later [20] — a problem that background-synthesis systems like Dreaming make more acute. None of the major providers have publicly addressed how they intend to handle persistent attack surfaces or the safety-review validity problem for continuously updating memory systems.