Version 5

OpenAI launched Dreaming V3 for ChatGPT on June 4, 2026 — a background memory synthesis system that periodically reviews conversation history and updates what it knows about users [1][2]. The Neuron published the first quantified accuracy metrics on June 7: ChatGPT's original memory had 41.5% factual recall accuracy in 2024; Dreaming V3 raises that to 82.8%, with preference adherence improving from 55.3% to 71.3% [5]. The 5x compute reduction that produced these gains also extends memory access to free users for the first time — accounting for earlier ambiguous reporting that Dreaming was rolling out broadly rather than only to Plus and Pro subscribers [5][6]. Competing approaches — Anuma's portable cross-model memory, Zep's agent tooling, Zaxy's coordinated architecture, and third-party tools like AgentMemory — continue developing in parallel.

The published baseline — wrong more than half the time in 2024 — shows how unreliable earlier AI memory implementations were, and The Neuron argues all major providers likely share similar accuracy problems without disclosing them [5]. The security attack surface grows with memory capability, documented by Palo Alto Networks and Sysdig as operational rather than theoretical, but no major provider has publicly addressed how they defend against prompt injection in continuously self-updating memory systems.

On June 4, 2026, OpenAI announced 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]. Unlike prior versions where users had to manage memory entries manually, Dreaming V3 periodically reviews conversation history and updates what the system knows about a user's preferences and goals. OpenAI added a user-facing memory summary page, giving subscribers visibility into what the system has retained [3][4].

The Neuron published the first quantified performance metrics for ChatGPT memory on June 7, 2026. OpenAI's original memory feature had a factual recall accuracy of only 41.5% in 2024 — wrong more than half the time on memory-dependent tasks. Dreaming V3 raises that to 82.8%, while preference adherence improved from 55.3% to 71.3% [5]. The 5x compute reduction that produced these gains also extended memory access to free users for the first time, not just Plus and Pro subscribers [5] — which accounts for The Verge's earlier reporting that the feature was 'rolling out to everyone' [6] rather than only paid tiers. The Neuron's framing of OpenAI's disclosure was notable: publishing poor baseline numbers rather than suppressing them was treated as unusual industry transparency, paired with a warning that all major AI assistants likely run similar accuracy problems without disclosing them [5].

The Dreaming launch sits in a broader field of competing approaches to AI memory. MIT MeMo research found a 26% LLM performance improvement when memory is kept architecturally separate from the base model [7]. Developer practitioners have documented compaction amnesia and context rot — where models lose track of prior decisions in complex workflows — and argued that structured selective memory retrieval is more reliable than extending context windows [8][9]. Anuma, building cross-model portable AI context, frames proprietary per-model memory as insufficient: its position is that user context should be portable across all AI models in a user-owned format rather than locked to a single platform [10][11]. Community voices have surfaced Zaxy's coordinated memory architecture [12] and Zep's persistent memory infrastructure for AI agents [13] as developer-facing alternatives, and a third-party AgentMemory tool appeared aimed at giving Claude Code persistent cross-session memory [14]. Practitioner frustration with re-explaining context to models each session continues [15].

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 [16]. Sysdig documented the first publicly known LLM-agent-driven post-exploitation chain in May 2026 [17], and SecurityWeek published a formal benchmark ranking the security posture of 100 AI agents [18]. A structural concern runs through practitioner commentary: Danny Livshits observed that a self-evolving agent voids its own safety review, since the version tested may not match the version running a month later [19] — a problem that background-synthesis systems like Dreaming make more concrete. None of the major providers have publicly addressed how they plan to defend against persistent attack surfaces or validate the safety of continuously updating memory systems.