Skip to main content

​
Token-aware Context Compaction

Agents with long conversations or large tool results can exceed the model’s context window. The ContextCompactor automatically manages context size before each LLM call.

​
Configuration

import { Agent, openai } from "@radaros/core";

const agent = new Agent({
  name: "long-conversation-bot",
  model: openai("gpt-4o"),
  contextCompactor: {
    maxContextTokens: 120_000, // for 128k context window
    reserveTokens: 4096,      // leave room for the response
    strategy: "hybrid",       // "trim" | "summarize" | "hybrid"
    summarizeModel: openai("gpt-4o-mini"), // cheap model for summaries
  },
});

​
Strategies

​
Trim

Drops oldest non-system messages first, keeping the system prompt and most recent exchanges intact.

​
Summarize

Uses a cheap model to summarize older messages into a single compact summary, preserving key context while reducing token count.

​
Hybrid

Trims first, then summarizes if still over budget. Best balance of speed and context preservation.

​
How It Works

The compactor hooks into the beforeLLMCall loop hook and runs before every LLM API call:
  1. Estimates token count of all messages
  2. If under budget, passes through unchanged
  3. If over budget, applies the configured strategy
  4. Returns the compacted messages to the LLM
System messages are always preserved. The most recent user/assistant exchanges are prioritized.