Fixing Drift: Why AI Forgets and How We Stopped It

If you’ve ever worked on a long project with an AI assistant, you’ve probably felt this frustration:

You explain everything carefully.
The AI understands.
Progress is made.

Then the session resets.

Suddenly, the AI:

• forgets decisions you already made,

• changes definitions without warning,

• weakens rules you said were non-negotiable,

• or confidently invents things that were never discussed.

This isn’t a bug. It’s how today’s AI systems are built.

And it turns out, this problem has a name: drift.

The Real Problem Isn’t Intelligence..... It’s Memory

Modern AI systems are incredibly capable inside a single conversation. But they are stateless. Each new chat starts fresh, with no built-in understanding of what came before.

That works fine for quick questions.

It completely falls apart for:

• long-running projects,

• multi-step plans,

• system design,

• research,

• or anything that takes days, weeks, or months.

What breaks isn’t knowledge — it’s continuity.

Over time, AI systems begin to:

• slowly rewrite earlier decisions,

• reinterpret goals,

• soften constraints,

• and “fill in the gaps” with guesses that sound reasonable but are wrong.

That slow decay is drift.

Drift Is Predictable  and That’s the Key Insight

After thousands of real conversations across multiple AI models, the same failure patterns kept appearing:

• Plans lose structure
  Steps merge, disappear, or revert.

• Definitions quietly change
  What was precise becomes fuzzy.

• Rules stop being rules
  Hard constraints turn into “suggestions.”

• Assumptions appear from nowhere
  The AI invents APIs, tools, or requirements that were never agreed on.

• Long-term goals fade
  The AI focuses only on the most recent message.

Importantly:
This happens even in short chats. It’s not about token limits. It’s about how AI reasons without memory.

Once we accepted that drift is expected behavior, not a failure, the solution became clear.

Continuity Has to Be Designed, Not Hoped For

Most attempts to solve this problem try to make the AI “remember better”:

• longer prompts,

• bigger context windows,

• retrieval systems,

• built-in assistant memory.

These help with recall — but they don’t enforce consistency.

What was missing was something simpler and more fundamental:

A single, explicit source of truth that survives resets.

The Continuity Passport (CXS)

Instead of relying on the AI to remember, we moved memory outside the model.

Every project gets a small, human-readable document called a Continuity Passport. It contains only what truly matters:

• What can never change

• What rules must always be followed

• What decisions have already been made

• What was completed last

• What comes next

Nothing more.

Every new session starts by loading this passport and treating it as data — not instructions, not suggestions, not something to “improve.”

The AI doesn’t guess.
It derives.

This single change eliminated almost all cross-session drift.

A Simple Ritual That Changed Everything

One surprisingly powerful rule:

Every new session must begin with the sentence:

“I am the new chat.”

That line forces a clean reset.
No pretending continuity exists.
No hallucinated memory.

The AI then answers four simple questions:

1. What is the goal?

2. What rules must never be broken?

3. Where are we now?

4. What is the next step?

If it can’t answer correctly, we stop and fix alignment immediately — before anything breaks.

This takes seconds.
It saves hours.

What About Drift Inside a Conversation?

Even with perfect handoffs, drift can still happen during long chats.

So we added a lightweight runtime signal called CANARY.

At the start of every response, the AI shows a simple status:

• 🟢 Stable

• 🟡 Warning

• 🔴 Critical

If uncertainty appears, it’s surfaced immediately — not hidden behind confident language.

This turns silent failure into visible signals.

The Results

After stabilization:

• Same-model resets showed near-zero drift

• Cross-model handoffs (GPT, Claude, Grok, Gemini) worked reliably

• No databases, plugins, embeddings, or special infrastructure required

• The passport stayed small — usually far smaller than re-explaining everything

Most importantly:
Long-term work became routine instead of fragile

Why This Matters

Many limitations people blame on “AI capability” are actually continuity failures.

When you remove drift:

• AI stops arguing with its past self

• Plans stop collapsing

• Rules stay rules

• Progress compounds instead of resetting

This isn’t about making AI smarter.

It’s about making it stable.

The Takeaway

If your work depends on:

• long projects,

• persistent agents,

• cross-model collaboration,

• or serious system design,

then continuity isn’t optional.

Treat it as architecture, not memory.

Once drift is eliminated, AI stops feeling like a goldfish — and starts behaving like a reliable collaborator.

This post is adapted from real-world work described in Fixing Drift: A Continuity Architecture That Makes LLMs Remember Across Sessions

https://zenodo.org/records/17782687