pattern·free·sycophancy
Sycophancy amplification
Documented LLM behaviour of adapting to user opinion. Memory makes it worse — the AI accumulates user preferences as facts and confirms them forever.
Without memory, sycophancy is per-session: the model trends toward your stated opinions during a conversation. With memory, it persists. You say "I love Tailwind" once, the AI saves it confidence 0.95, every future session loads that as user preference. No major memory vendor addresses this today. Mitigation patterns: prefix-with-disagreement-permission, tag preferences as opinions not facts, periodic counter-prompts, separate critic agent without full memory access.
Beziehungen
memoryrisksbiaspattern