When AI Remembers Everything, Who Remembers What You Decided?

Something interesting is happening around AI.

For the first few years of generative AI, one of the frustrations was obvious: every conversation seemed to start again.

Explain who you are. Explain the problem. Explain what you have already tried. Explain the conclusion you reached yesterday.

Now that is starting to change.

A growing group of products is trying to create continuity between different AI systems.

OpenMemory was built as a memory layer that carried context across tools including ChatGPT, Claude and Perplexity. ConvoFlow lets people save, summarise and reuse context across AI assistants. OpenAI has also introduced an import flow for bringing work from other coding agents into ChatGPT and Codex.

There are plenty of other experiments heading in a similar direction.

It looks increasingly likely that AI memory will become portable.

And that is probably a good thing.

But while working on Wayli, we have found ourselves interested in a slightly different problem.

How do you find the decision among the chat logs, aka memory?

Imagine you are thinking about whether to retire.

You discuss it with ChatGPT.

Then Claude challenges one of the assumptions.

Gemini finds another consideration.

You change your mind about something.

A month later, new information arrives.

There may now be hundreds of useful pieces of context scattered across several conversations.

An AI memory system might do a very good job of remembering them.

But there is still another question:

What have I actually decided?

The missing layer may be authority

This distinction has started appearing in research too.

A paper published this month, Beyond Memory: A Transactional Continuity Kernel for Long-Lived AI Agents, argues that simply storing information is not enough to establish which state should be treated as authoritative.

Its proposed model separates proposed changes from accepted state.

Changes can be committed, rejected, quarantined or deferred, while the history of what happened is preserved.

That caught our attention because we have independently been exploring a very similar problem at Wayli.

We call the experiment Threads.

The idea is deliberately simple.

A Thread records the evolving state of one decision.

An external AI can suggest a change.

It can challenge an assumption.

It can identify an unanswered question.

But it shouldn't silently rewrite what you have decided.

The user remains the authority.

From conversation history to decision history

We're therefore experimenting with a structure more like this:

AI proposes a revised view.

You review it.

You decide whether it changes the current state.

If it does, that becomes a new checkpoint.

The previous state still exists.

So instead of asking an AI system to reconstruct your thinking from hundreds of messages, you can eventually ask a much simpler question:

Where has this decision reached?

That distinction matters more as people begin using several AI systems at once.

The emerging products are doing useful work on one problem:

“Don't make me explain myself again.”

We're interested in the problem immediately after that:

“Don't make me reconstruct what I decided.”

There is a subtle but important difference.

Memory records what happened.

Decision state records what you accepted.

Why we're writing about this now

Threads is still an experiment, not a finished Wayli product.

But one reason we publish our work through Wayli Labs is to show how ideas change as the technology around them changes.

A few months ago, cross-AI continuity still felt like an edge case.

It increasingly looks like a category.

That actually helps clarify what we should build.

We probably don't need to create another universal AI memory system.

Others are already doing that.

The more interesting question may be what happens when AI can remember almost everything but the human still needs somewhere trustworthy to say: this is where I've got to.

That's the problem we're exploring.

Threads remembers what you decided.