OpenClaw Wants Your ChatGPT History Inside the Memory Loop

Every assistant vendor says they want to remember you. Usually that means one of two bad experiences: either the system forgets everything the moment the tab closes, or it remembers in ways that feel creepy, opaque, and impossible to correct. OpenClaw’s latest memory push is interesting because it is trying to solve a real user problem, but it is doing so by dragging old chat history into one of the most privileged loops in the product. That is where useful product design stops being a UX feature and starts becoming storage engineering.

In OpenClaw 2026.4.11, the headline move is not another channel integration or model option. It is the addition of ChatGPT import ingestion into Dreaming and memory-wiki, plus two new diary surfaces called Imported Insights and Memory Palace. The release notes say Dreaming can now inspect imported source chats, compiled wiki pages, and full source pages directly from the UI. Translated out of changelog-speak, OpenClaw wants your assistant to treat your prior ChatGPT life as usable working context instead of dead archive material.

That is the right product instinct. Real users do not begin with a blank slate. They have months or years of conversations scattered across ChatGPT exports, notes, docs, screenshots, and half-remembered threads. A memory system that only starts collecting after installation is technically neat and practically late. OpenClaw is trying to shorten that gap. If the platform can ingest historical conversations and surface them through a UI that is actually inspectable, then memory becomes less of a gimmick and more of a migration story.

But there is a catch, and it is not a small one. The moment imported chat exports begin feeding Dreaming and memory-wiki, the product is no longer just retrieving memory. It is ingesting untrusted user-supplied archives, parsing timestamps and paths, creating rollback records, exposing page views, and potentially replaying imported material into later retrieval flows. That is why the security analysis on the related PR is so revealing. The flagged issues include path traversal risks around import run IDs and digest page paths, denial-of-service risks around malformed timestamps and unbounded page reads, and leakage concerns around filesystem paths returned to operator.read clients. Whether every finding lands exactly as written matters less than the broad truth it exposes: imported memory is an attack surface.

Memory import changes the trust model

The Active Memory docs already framed memory as a plugin-owned blocking sub-agent that can run before the main reply on eligible persistent chats. Safe defaults are explicit: direct-message style sessions, recent-mode queries, balanced prompt style, 15-second timeout, 220-character summaries, transcript persistence off. Those defaults are sensible because proactive memory is powerful enough to make mistakes faster than reactive memory does. Add imported ChatGPT archives to that loop and the trust model gets harder again.

This is the product tradeoff most assistant companies keep trying to skip. Users absolutely want continuity. They do not want to manually tell the system what it should already know. But continuity without provenance is how assistants become overconfident liars. If a memory system can pull from imported chat logs, compiled wiki pages, and full source pages, then the interface has to help users answer basic questions: where did that recollection come from, why was it selected now, is it stable preference or stale artifact, and how do I turn it off when it gets weird?

OpenClaw is at least moving in a more honest direction than many competitors. Exposing imported material through Imported Insights and Memory Palace acknowledges that memory should be inspectable. That sounds obvious, but much of the industry is still shipping “personalization” as invisible prompt spaghetti. Visibility is not a nice extra here. It is the precondition for trust.

The useful future is bounded, not magical

There is also a more strategic point buried in this release. Agent products are converging on a world where context is assembled from multiple stores: active transcript, long-term notes, imported history, maybe shared team docs, maybe external wiki pages, maybe a calendar or inbox. The winning products will not be the ones that remember the most. They will be the ones that can compose those sources without turning relevance into chaos.

That is why the specifics in the docs matter. OpenClaw’s Active Memory guidance pushes bounded recall and explicit gating instead of “memory everywhere.” It runs only for eligible persistent chats. It can be switched off per session. It has prompt styles that change recall eagerness. It can skip the turn entirely if no model resolves or the connection is weak. Those are the right instincts because the product problem is not merely to remember more. It is to remember without hijacking the conversation.

For developers building on top of systems like this, the action items are pretty clear. First, treat imported context as data ingestion, not as a cute enhancement. Validate paths, clamp file and line reads, and assume archives can be malformed. Second, keep memory debug visibility on while tuning and keep recall scopes narrow at first. If the product cannot explain its own memory behavior during testing, it definitely will not explain it in production. Third, distinguish between preferences and facts. Users often want stable tastes, recurring habits, and ongoing project context remembered. They usually do not want one offhand joke from six months ago to become policy.

There is also an organizational lesson here. Teams adopting agent memory need governance before they need cleverer prompts. Decide which imported sources are allowed, who can inspect them, how retention works, and how users correct or delete bad memory. Imported context sounds like a UX improvement, and it is, but it also creates compliance, security, and support questions the minute the feature becomes successful.

My read is that OpenClaw is aiming at the correct target. Personal and team assistants should be able to inherit useful historical context instead of pretending history began at install time. But this is one of those features where ambition and restraint have to grow together. The more an assistant knows, the more the product has to prove it knows how it knows. OpenClaw’s new memory surfaces are promising because they move in that direction. The real test is whether the project keeps treating memory import as core platform engineering rather than a flashy retention trick.

Sources: OpenClaw release v2026.4.11, PR #64505, OpenClaw Active Memory docs