Skip to main content

​
Overview

Context pollution is the #1 production failure mode. Old/failed messages accumulate, causing agents to drift toward “plausible-but-wrong” outputs. The ContextCurator automatically cleans conversation context before each LLM call.

​
Quick Start

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

const agent = new Agent({
  name: "clean-context-agent",
  model: openai("gpt-4o"),
  contextCurator: {
    enabled: true,
    failedResultHandling: "deprioritize",
    relevanceDecay: {
      enabled: true,
      halfLifeTurns: 10,
      minWeight: 0.1,
    },
    maxFailedResults: 1,
  },
});

​
Failed Result Handling

Tool calls fail. Those error messages stay in context and confuse the model. The curator handles them:
StrategyBehavior
"remove"Remove failed results entirely (keep last N)
"deprioritize"Prefix with [PREVIOUS ERROR - may not be relevant]
"summarize"Compress to one-line summary

​
What counts as a “failed result”?

The curator detects errors by pattern matching:
  • [ERROR] markers
  • Error: prefixes
  • HTTP 4xx/5xx status codes
  • timeout, ECONNREFUSED, Permission denied
  • Stack traces

​
Relevance Decay

Messages far back in the conversation become less relevant. The curator assigns a weight based on turn distance: 
weight = max(minWeight, 0.5 ^ (turnDistance / halfLifeTurns))
Messages below the threshold are either dropped or compressed to one-line summaries. Messages containing entities mentioned in the current query are always preserved.

​
Parameters

ParameterDefaultDescription
halfLifeTurns10Turns until weight halves
minWeight0.1Floor weight (below this → consider dropping)

​
Clean-Room Mode

Instead of filtering the full conversation, build a purpose-specific context: 
const agent = new Agent({
  contextCurator: {
    enabled: true,
    cleanRoomMode: true, // Build filtered view per call
  },
});
Clean-room context includes:
  1. System prompt
  2. Memory context
  3. Messages containing entities from the current query
  4. Last N recent messages

​
Standalone Usage

import { ContextCurator } from "@radaros/core";

const curator = new ContextCurator({
  enabled: true,
  failedResultHandling: "summarize",
  relevanceDecay: { enabled: true, halfLifeTurns: 8, minWeight: 0.15 },
  cleanRoomMode: false,
});

const curated = curator.curate(messages, "What are our Q4 results?");

​
Events

EventPayload
context.curated{ runId, originalCount, curatedCount, failedRemoved }