← Retour au blog
tech 31 July 2026

The Session You Cannot Take With You

Inference APIs promised simplicity: an input, an output. But the reality of session portability is quite different. Let's explore why this matters for you and your project.

Article inspired by the original source
The session you cannot take with you ↗ earendil.com

Introduction

Inference APIs have revolutionized how we interact with AI, offering a simple interface: an input and an output. However, behind this apparent simplicity lies an increasing complexity, especially concerning session portability. Why is this an issue? And why should you care?

The Original Promise

Initially, the idea was simple: send a request, receive a response, and keep the session history. This would allow data reuse, interaction analysis, and even transferring this session to another model. In theory, it sounds ideal. However, the reality is different. Prompt caches live on someone else's GPUs, tokenization varies between models, and sampling is intentionally non-reproducible.

The Shift Towards Non-Portability

Today, we observe a worrying trend: inference APIs increasingly return a mix of text and provider-bound state, often non-portable. For example:

  • Reasoning tokens: billed to the user but returned as encrypted blobs.
  • Web searches: the model sees sources, but the client does not.
  • Compacted context: only decryptable by the original provider.

This evolution means that the session history on your machine is just a partial view, whose operational state belongs to the inference provider.

Why It Matters

For a developer or entrepreneur, understanding this limitation is crucial. It affects how you design your AI-based products and services. Non-portable sessions can limit your ability to switch providers or integrate different models. In 2023, 67% of companies using AI expressed concerns about over-reliance on a single provider, according to a Gartner study.

A Practical Solution

Ideally, session portability should allow data transfer from one provider to another without information loss. An example of this concept could be:

``javascript const transcript = session.export(); revokeCredentials(oldProvider); session = newProvider.continueFrom(transcript); ``

The transcript must contain enough intelligible information for another model to continue the work. This requires collaborative efforts between providers to establish interoperability standards.

Conclusion

Session portability is not just a luxury; it is a necessity for the evolution of AI technologies. As a tech decision-maker, you need to be aware of these issues to avoid being trapped in a closed ecosystem.

Let's discuss your project in 15 minutes.

API d'inférence portabilité des sessions interopérabilité IA fournisseurs IA dépendance technologique
Deepthix newsletter · 100% AI · every Monday 8am

An AI agent reads tech for you.

Our AI agent scans ~200 sources per week and ships the best articles to your inbox Monday 8am. Free. One click to unsubscribe.

Visit the newsletter page →

Want to automate your operations?

Let's talk about your project in 15 minutes.

Book a call