Version 1

OpenAI launched 'Dreaming' (Dreaming V3) for ChatGPT on June 4, 2026 — a background synthesis system that continuously updates user memory from conversation history rather than storing static snapshots [1][2][6]. The rollout starts with Plus and Pro subscribers. Concurrently, startup Anuma is positioning as a cross-model portable memory layer, arguing that proprietary per-model memory forces users to manually carry context between AI systems [10][9]. Academic research from MIT (MeMo) reports a 26% performance gain from keeping memory architecturally separate from the base model [11], and Palo Alto Networks has documented prompt injection as a live attack vector against persistent AI memory [12].

How AI assistants handle persistent memory is becoming a key axis of platform differentiation. OpenAI's Dreaming locks richer context into ChatGPT specifically; Anuma's portable model treats memory as user-owned infrastructure spanning all models. The two approaches have different implications for user lock-in, privacy, and the economics of AI platform competition. The concurrent security research makes the stakes concrete: more capable proactive memory systems also expand the attack surface.

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]. The system rolls out first to Plus and Pro subscribers. OpenAI's blog frames it as a move from passive storage to proactive synthesis: the model periodically reviews conversation history and updates what it knows about a user's preferences and goals [1]. The announcement drew rapid social amplification within hours, with dozens of retweet threads and commentary posts [3][4][5][6].

The Dreaming launch sits in a broader context of documented practitioner frustration with context loss. A Hacker News discussion from late May described 'compaction amnesia and context rot' — the pattern where AI coding tools like Codex progressively lose track of prior decisions when handling complex, multi-step workflows [7]. Earlier developer forum posts proposed 'memory-first conversational architecture' as more reliable than extending context windows, arguing that selective structured memory retrieval beats raw context length [8]. These practitioner pain points frame the commercial urgency behind OpenAI's memory investment.

A competing architectural philosophy has emerged from Anuma, a startup building cross-model portable AI context [9]. Rohan Paul described the core problem plainly: 'Most AI workflows break because the user has to carry the context manually' [10]. Anuma stores context, preferences, and goals in a format portable across ChatGPT, Claude, and other systems, positioning itself as user-owned infrastructure rather than platform-locked memory. This directly contrasts with OpenAI's proprietary approach — if portable memory layers gain traction, they could reduce the stickiness of model-specific memory features.

Two additional technical threads matter. MIT's MeMo system, circulated in late May, reports a 26% LLM performance improvement by keeping memory architecturally separate from the base model, with the memory module updatable without retraining [11]. Separately, Palo Alto Networks documented that indirect prompt injection can poison AI long-term memory, causing persistent behavioral changes across sessions [12]. The combination of more proactive, trusted memory systems and this known attack vector creates a security tension that none of the major providers have publicly addressed.