Persona Drift — Catching It Across Turns

~14 min · conversation, persona, drift

Level 0Apprentice

0 XP0/100 lessons0/14 achievements

0/120 XP to next level120 XP to go0% complete

The persona that erodes

By turn 20, your warm, terse support agent has become a chatty, hedging assistant who apologizes too much. The model isn't broken — it's mirroring the user's longer messages, hedging because it saw hedging earlier in context, and slowly forgetting the system prompt's tone clauses under the weight of the conversation.

Three drift mechanisms

Mirroring — model adopts the user's tone, length, and vocabulary.
Self-imitation — once it hedges once, hedging is in context, so it hedges more.
System-prompt dilution — system is 200 tokens; turn-50 history is 20,000 tokens; the system's relative weight drops.

Counter-drift tactics

Re-anchor periodically (every N turns, re-inject a one-line tone reminder).
Use compaction to remove drift-y assistant turns from history.
Add explicit anti-drift sentences ("Do not adopt the user's tone").
Run a tone classifier on assistant outputs and alert on drift.

Code

Periodic re-anchor·python

if turn_count % 10 == 0:
    messages.append({
        "role": "user",
        "content": "[system reminder] Stay terse. Open with the next step. Do not apologize."
    })

External links

Exercise

Capture 100 long conversations from your system. Score the assistant outputs for tone drift (terseness, hedging frequency). Plot drift over turn number. If real, ship a re-anchor at the half-life.

Progress

Progress is local-only — sign in to sync across devices.

← PreviousMemory — When the Model Should Remember and When It Shouldn't Next →Long Conversations — Compaction Strategies

Spotted a bug or have feedback on this page?Report an Issue

Comments 0

🔔 Reply notifications (sign in)

No comments yet — be the first.