Backboard homeBlog

Rob Imbeault

·

Adaptive Context Windows Across 17,000+ Models

Backboard now includes Adaptive Context Management, a system that automatically manages conversation state when your application moves between models with different context window sizes.

With access to 17,000+ LLMs on the platform, model switching is common. But context limits vary widely across models. What fits in one model may overflow another.

Until now, developers had to handle that manually. Adaptive Context Management removes that burden. And it's included for free with Backboard.


The Problem: Context Windows Are Inconsistent

Different models support different context window sizes. If an application starts a session on a large-context model and later routes a request to a smaller one, the total state can exceed what the new model can handle.

That state typically includes:


  • system prompts

  • recent conversation turns

  • tool calls and tool responses

  • RAG context

  • web search results

  • runtime metadata


Most platforms leave this responsibility to developers. In multi-model systems, that quickly becomes fragile.


Diagram of Adaptive Context Window Management: a 16K-token context in Model A (system prompts, recent messages, tool calls, RAG context, web search results, runtime metadata) transitions to an 8K-token context in Model B, where 20% is reserved for raw state and 80% holds an intelligently summarized compressed state, with a usage widget showing 5,600 of 8,000 tokens used.


Introducing Adaptive Context Management

Backboard now automatically handles context transitions when models change. The system works as follows:


  • 20% of the model's context window is reserved for raw state

  • 80% is freed through intelligent summarization


Within the 20% budget, Backboard prioritizes:


  • system prompt

  • recent messages

  • tool calls

  • RAG results

  • web search context


Whatever fits inside this budget is passed directly to the model. Everything else is compressed.


Intelligent Summarization

When compression is required, Backboard summarizes the remaining conversation automatically:


  1. First we attempt summarization using the model the user is switching to.

  2. If the summary still cannot fit, we fall back to the larger model previously in use to generate a more efficient summary.


You Should Rarely Hit 100% Context Again

Because Adaptive Context Management runs continuously during requests and tool calls, the system proactively reshapes the state before a context window is exhausted.


Developers Can See Exactly What Is Happening

We expose context usage directly in the msg endpoint:

"context_usage": { "used_tokens": 1302, "context_limit": 8191, "percent": 19.9, "summary_tokens": 0, "model": "gpt-4" }


The Bigger Idea

Backboard was designed so developers can treat models as interchangeable infrastructure. Adaptive Context Management is another step toward that goal. Applications can move freely across thousands of models while Backboard ensures the conversation state always fits.

Adaptive Context Management is available today through the Backboard API. Start building at docs.backboard.io

Backboard

Ottawa, Canada
Burlingame, California

Backboard © 2026