Conversation Repetition

Goal Conversation Quality Frequency Very Common Category Agent Interaction Published Jul 27, 2026 View source on GitHub ↗

Issue

The agent restates information, asks a question, or repeats an instruction that it (or the user) already covered earlier in the same conversation, as if encountering it for the first time. This differs from coherence loss in scope: repetition is the narrow, directly observable symptom — the same content appearing twice — whereas coherence loss is the broader state-tracking failure that often, but not always, produces it (a single instance of repetition can also come from a template default firing regardless of history).

Frequency: Very Common

Symptoms

Agent asks a question whose answer was already given earlier in the conversation
Agent re-states a fact, caveat, or instruction verbatim or near-verbatim from an earlier turn
Boilerplate disclaimers or setup text repeat every turn even after being acknowledged once
User responses shift toward “I already told you…” or “you just said that”
Repeated content adds no new information relative to the first occurrence

Root Cause

Repetition typically comes from one of two mechanisms: either the model regenerates content from a static template or system instruction on every turn without checking whether it was already said (e.g. a standard disclaimer re-appended each response regardless of prior turns), or the model fails to retrieve/weight an earlier turn’s content strongly enough during generation and independently re-derives the same question or statement as if starting fresh. Both share the same missing safeguard — there’s no check comparing candidate output against what has already been said in the conversation before finalizing a response.

Example

Turn 2:  User: "My account email is jsmith@example.com."
Turn 3:  Agent: "Got it, using jsmith@example.com."

Turn 9:  Agent: "Before I proceed, could you confirm the email
         associated with your account?"

Turn 10: User: "I gave you that seven turns ago — it's
         jsmith@example.com."

Turn 14: Agent: "Just to note, refunds typically take 5-7 business
         days to process."
Turn 19: Agent, discussing a different aspect of the same refund:
         "Just to note, refunds typically take 5-7 business days to
         process." (identical caveat, no new context justifying the
         repeat)

Turn 20: User: "You've told me that twice now, can we move on?"

Statistics

Finding	Context
An estimated 10-20% of sessions exceeding 15 turns contain at least one instance of the agent asking for information already provided	Typical range across production conversational agent logs
Repeated boilerplate/disclaimer text appears in a majority of multi-turn sessions using static per-turn templates without a “already stated” check	Estimated from template-based agent deployments
Adding a said-before check before finalizing a response reduces flagged repetition incidents substantially	Reported range across teams that added conversation-content deduplication

Mitigations

Said-before check: Before finalizing a response, compare candidate content (questions, facts, disclaimers) against a record of what has already been stated in the conversation, and suppress or rephrase exact/near-duplicates.
Slot-filled memory for collected inputs: Once a piece of information (email, account ID, preference) is provided, store it in structured session state so the agent reads from state rather than re-asking.
One-time boilerplate: Mark template disclaimers/setup text as “state once per session” rather than “include every response,” with an explicit trigger if circumstances change enough to warrant a repeat.
Repetition-aware response generation: Include a compact summary of already-stated key facts and questions directly in the prompt context so the model has an explicit signal of what not to repeat.
User-flagged repetition fast path: When a user says “you already told me” or similar, treat it as ground truth immediately rather than re-verifying, and suppress that content for the rest of the session.

Production Signals

Key Metrics

Metric	Description	Alert Threshold
repeated_question_rate	Share of sessions where the agent asks for already-provided information	Alert if > 10%
duplicate_content_rate	Share of turns containing near-duplicate text of an earlier turn in the same session	Alert if > 15%
user_repetition_complaint_rate	Rate of explicit user statements like “I already told you”	Alert if > 5%

Alerts

Alert	Condition	Severity	Response
Repeated request for provided information	Agent asks for a value already present in session state	Medium	Trigger state-lookup fix, log for slot-filling review
User flags repetition	Explicit “already told you” style user statement detected	Low	Log session, review said-before check coverage

Conversation Coherence Loss - a frequent root cause of repetition, since forgotten state leads directly to re-asking or re-stating
Conversation Length Explosion - repeated content inflates turn count without adding information, contributing directly to unbounded conversation growth
Over-Clarification - repeated questions are a specific case of over-clarification when the repeated content is itself a clarifying question