AI Persistent Memory: ChatGPT Dreaming and the Cross-Session Context Race

What

Why it matters

Open questions

• Other AI assistant providers likely have memory accuracy problems comparable to ChatGPT's 41.5% baseline [5] — will any publish similar benchmarks, or does OpenAI's disclosure remain the exception?

• Does Dreaming V3's 82.8% recall accuracy hold as user histories grow longer and more complex, or does performance degrade at scale? [5][21]

• Can local, on-device memory tools like Kocoro [10][13] gain enough traction to serve as privacy-preserving alternatives, or do they lack the distribution and model integration that platform-native solutions provide?

• A self-evolving agent continuously updates its own state, so the version that passed safety review may not match the version running weeks later [19] — how do providers plan to handle this for background-synthesis systems like Dreaming?

Narrative

Timeline

• 2026-05-10: Sysdig documents the first publicly known LLM-agent-driven post-exploitation chain. [17]
• 2026-05-26: Hacker News post documents compaction amnesia and context rot in Codex on complex multi-step workflows. [8]
• 2026-05-30: MIT MeMo research claims 26% LLM performance gain from memory kept architecturally separate from the base model. [7]
• 2026-06-02: Danny Livshits observes that a self-evolving agent voids its own safety review, as the tested version may not match the production version. [19]
• 2026-06-03: SecurityWeek publishes a benchmark ranking the security posture of 100 AI agents. [18]
• 2026-06-04: OpenAI announces Dreaming V3 for ChatGPT — background memory synthesis with a user-facing summary page — rolling out initially to Plus and Pro users. [1][2][3][4]
• 2026-06-04: Anuma cross-model portable memory discussed as a structural alternative to proprietary model-specific memory. [11][12]
• 2026-06-04: Palo Alto Networks research on indirect prompt injection poisoning AI long-term memory surfaces in context of the Dreaming launch. [16]
• 2026-06-05: The Verge reports Dreaming rolling out to everyone, suggesting broader availability than the Plus/Pro-only initial framing. [6]
• 2026-06-05: Third-party AgentMemory tool appears to give Claude Code persistent cross-session memory. [20]
• 2026-06-06: Developer commentary attributes roughly 5x lower compute cost to Dreaming V3, adding specificity to OpenAI's vague 'compute-efficient' claim. [21][24]
• 2026-06-07: The Neuron publishes quantified memory accuracy metrics: 41.5% factual recall in 2024 rising to 82.8% with Dreaming V3; preference adherence up from 55.3% to 71.3%; 5x compute reduction extends memory to free users. [5]
• 2026-06-08: Kocoro promoted as an open-source Mac-native AI agent running local background memory review of past sessions, explicitly framed as an alternative to context-window scaling. [10][13]
• 2026-06-09: Alibaba Cloud opens a MemoryAgent Arena track at its Qwen Cloud Global AI Hackathon, signaling institutional competition around agent memory. [15]
• 2026-06-09: RecallOS announces its first chapter as a persistent-memory product. [14]

Perspectives

Tensions

• OpenAI published its memory accuracy baseline (41.5% recall in 2024); The Neuron argues all major AI providers likely share similar accuracy problems but have not disclosed them, framing OpenAI's transparency as the exception rather than the norm. [5]
• OpenAI's Dreaming locks richer memory into ChatGPT specifically; Anuma, Kocoro, Zaxy, and Zep argue memory should be portable, user-owned, or locally controlled rather than platform-managed. [1][11][12][23][22][10][13]
• Dynamic background synthesis (Dreaming) vs. static explicit memory: practitioners argue proactive synthesis may introduce context drift as a new failure mode rather than solving reliability. [1][9]
• More capable persistent memory expands the prompt-injection attack surface; Palo Alto Networks and Sysdig confirm this is operational, but neither OpenAI nor other providers have publicly addressed how they will defend against it. [16][1][17]
• Self-evolving agents that continuously update their own memory may void prior safety reviews; Danny Livshits argues the tested version is not the version running a month later — a structural problem that background-synthesis systems like Dreaming make concrete. [19][1]