OpenAI launches an active "dreaming" memory system for ChatGPT to continuously update user preferences across conversations

eh looks helpful, but feels like its being over-marketed. Don't feel comfortable setting the bar to RAG-eque related curation as "memory". and the weird comparisons to 'dreaming' and consolidation bothers me. Not a good thing to condone it as anything more than better externalized context management: summaries, retrieval, ranking, freshness, and maybe policies trained to decide what should be saved or surfaced. (Because once we let every retrieval/summary system get branded as 'memory' or 'dreaming', the terms stop pointing at the harder problems: persistent state, actual learning, agency, continuity, and what the system is really doing. bleh )

That’s useful though! But it’s not the same as persistent latent state, ongoing self/world/task-state, or actual continuous learning.

Calling it memory is already a stretch. Calling it dreaming makes the stretch also do a backflip. >.>'

personally i'd like to see us move away from externally managed summaries/ranking systems and closer to Stateful model arch around the AR core, and real memory/latent state research stuffs please T_T 🥺🙏

Fair enough

We are pretty early, and there’s an extremely limited number of people impactfully innovating useful systems atop of Codex/frontier-agents

Plus, for casual users, all a proper compaction/hydration (pruning) system would do is: eat more tokens/usage. (Until more efficient methods are discovered)

So, all in all, since we’re operating in frontier territory, there’s a shared responsibility of building your own solutions

We’ve seen this exemplified by the slew of engineers/products OpenAI/Codex have recently acquired.. 2-3 of them I was following before they got hired by them

It’s easier than ever to participate in impactful frontier agentic engineering innovation

Just gotta R&D

Fair point. That’s actually the direction I moved in.

The screenshot you shared is basically: durable artifacts, lane ownership, scout reports, post-compaction checks, and reconstruction from disk state before continuing.

I did something similar, but split it into an operational contract instead of one big prompt.

My workaround now has three modes:

Audit mode No code changes. No deploys. Only recover state, reconcile docs, identify risks, and decide the next phase. Controlled apply mode Small microphases only: patch, lint, hash check, backup, controlled SFTP, remote UAT, evidence, checkpoint.

Micro-UI mode One visual change only. No backend. No SQL. No SFTP. If it gets complex, it hands control back to the stronger model. The important part is the compaction circuit breaker.

After auto-compaction, I no longer trust the previous execution contract. The agent must stop: no commands, no patches, no deploys, no UAT. First it has to recover the minimal durable state, confirm the objective, and resume from checkpoint. So yes, we can be more creative.

But the fact that we need to build this ourselves is the issue. Serious agentic workflows need native compaction-aware execution: transparent context accounting, durable checkpoints, artifact continuity, and a hard safety boundary when the model state has been compressed.

Thanks and best. 🙌