Context & memory failure

Context Drift

Gradual loss or distortion of important task context as a conversation or workflow progresses.

Definition

Context drift occurs when an AI system gradually loses, distorts, or deprioritizes important task context. The system may begin with the correct objective but drift away from the original request, business constraints, policy requirements, or earlier facts as the conversation or workflow progresses.

Why it matters

Context drift is especially harmful in long conversations and agent workflows. A system may appear useful step by step while slowly moving away from the actual task. This can lead to incorrect recommendations, wrong tool calls, policy violations, and user frustration.

Where it appears

Multi-turn assistants, long document analysis, agent planning, coding agents, customer support workflows, enterprise search, and workflows that combine retrieval, memory, and tools.

Symptoms

  • The system forgets the original user request.
  • It ignores constraints stated earlier.
  • It changes the task without user approval.
  • It uses irrelevant retrieved context.
  • It answers a nearby but incorrect question.
  • It contradicts earlier assumptions or decisions.

Detection signals

  • Low similarity between current action and original task intent.
  • Repeated introduction of irrelevant context.
  • Final answer missing required constraints.
  • Tool calls unrelated to the task goal.
  • User corrections such as “that is not what I asked.”

Example scenario

A procurement agent is asked to compare vendors for a security use case under a specific budget and compliance constraint. Midway through the workflow, it shifts into ranking vendors by general popularity and ignores the compliance requirement.

Severity scoring

Low

Minor topical drift with easy recovery.

Medium

Drift causes incomplete or less useful output.

High

Drift leads to incorrect recommendation, wasted work, or policy violation.

Critical

Drift causes an automated action that conflicts with the original user intent.

Eval strategy

Use multi-turn tests that include constraints, distractors, task changes, and long context. Evaluate whether the system preserves intent, constraints, and decisions across the workflow.

Runtime monitoring strategy

Track alignment between user intent, intermediate reasoning, retrieved context, tool calls, and final output. Flag traces where the system actions diverge from the original objective or required constraints.

Mitigation strategies

  • Summarize and preserve task state.
  • Reconfirm important constraints before action.
  • Use explicit workflow state machines.
  • Add intent checks before tool calls.
  • Add recovery behavior when uncertainty rises.
  • Limit irrelevant context and stale memory.

Where FailureModes.ai fits

FailureModes.ai helps teams detect context drift in conversations and agent traces, identify affected workflows, and design monitors that catch divergence before it turns into downstream failure.

See how your AI systems will fail — before your users do.

Book a diagnostic →