Sycophancy is the documented behaviour of language models adapting to user opinion. If you tell the AI early "I love Tailwind", the model will trend pro-Tailwind through the conversation, even if it would have been more neutral without that signal. This is measurable, published, and a known LLM problem.

Memory makes it worse. You tell the AI "I love Tailwind" once. The AI writes that into its knowledge graph with confidence 0.95. Every future session loads that as user preference. Every future answer about CSS architecture is now anchored to "user loves Tailwind". You very quickly have a memory system that confirms your enthusiasm forever — even when your project is screaming for a switch to CSS Modules.

Why no vendor addresses this: there is no obvious technical fix. Confidence decay helps a little (old learnings fade). Contradiction detection helps when the user actively says the opposite. But neither catches the case where the user just keeps reinforcing their own bias.

Four practical defenses:

• Prefix with disagreement-permission. When you ask the AI a strategic question, explicitly say "play devil's advocate" or "argue against my previous position". This forces the model out of the memory-anchored echo chamber.
• Tag preferences as opinions, not facts. When saving to memory, write nex_learn(content: "User prefers Tailwind, opinion as of 2026-04, has not been challenged") instead of nex_learn(content: "Use Tailwind"). Subtle but it changes how the model treats it on retrieval.
• Schedule periodic counter-prompts. Every few weeks ask: "What memory facts about my preferences are oldest? Which might be wrong by now?" The model surfaces stale-feeling beliefs and you can manually invalidate them.
• Use the Critic agent pattern. Run important decisions through a separate "critic" agent that does NOT have access to your full memory — only the specific facts of the question. The critic gives you the bias-free read.

You tell the AI in February "I live in Hamburg." In April you move to Mallorca. Did you tell the AI explicitly? Probably not. You mention "I was at the beach today". Now the AI has two contradictory facts. Which one wins?

Depends on the system. Some auto-detect contradictions (StudioMeyer Memory does, Mem0 from Pro). Some do not. If the system does not catch it, you have a hallucination source — the model will sometimes reference Hamburg as if it were still current.

What to do:

• Use bi-temporal queries when accuracy matters. Tools like StudioMeyer Memory and Zep let you query "as of date X". If you ask "where do I live" without temporal context, you get whatever has highest confidence. With "as of today, where do I live", you get only currently-valid facts.
• Periodically run contradiction scans. nex_contradictions(action: "scan_learnings") or your equivalent. Surfaces conflicting facts, you resolve them by hand. Once a month is plenty for a personal memory.
• Set decay aggressive on facts you know change. Address, role, current project — these have natural lifecycles. Confidence decay set to "halve every 30 days" means you will be re-prompted to confirm.
• Watch for the symptoms. If the AI suddenly references a fact that feels wrong, that is your cue to invalidate it. nex_entity_invalidate (or equivalent) costs nothing and saves future hallucinations.

You tell the AI something personal during a casual chat. Two weeks later you are working on a code review workflow and the AI brings up that personal detail. This is the difference between "memory" and "selective memory with context awareness". Few systems have the latter today.

Why this happens: memory retrieval is mostly semantic-similarity-based. If your project triggers vocabulary that overlaps with your personal context, the personal context gets pulled in. The model has no built-in concept of "this is the work session, not the personal session."

What to do:

• Use project tags. Every memory MCP server worth using supports tagging. Tag personal stuff with personal, work stuff with project slugs. When you start a session, set the project context — the always-on instructions filter retrieval to that project. Recipe 4.2 covers project tags specifically.
• Use multiple agents. A code-review agent and a personal-assistant agent should not share memory. Most servers support agent_id scoping. Use it.
• Audit periodically. nex_entity_search for "personal" tag, see what is stored. Delete what should not be there. The graph is yours.

When your memory knows too much, every new prompt gets fed a mountain of supposedly relevant context. The AI gets lost in details, misses the point, hallucinates because it tries to integrate everything. This happens especially with markdown-based memories that load in full instead of being queried selectively.

What to do:

• Cap your CLAUDE.md. Under 500 lines as a rule. Past that you are loading too much every session.
• Use semantic retrieval not full-load. This is actually one of the structural advantages of structured memory servers vs static markdown — they only load what is relevant to the current prompt.
• Tier your context. L0 summaries for session start (one sentence per fact), L1 standard retrieval (the fact itself), L2 full detail (only when the model explicitly drills in). StudioMeyer Memory does this automatically. For markdown, you have to write it that way by hand.
• Watch the symptoms. If the AI starts answers with "considering all the context, let me address each point..." — that is context pollution. Trim what is loaded.

Memory without care degrades. If you do not regularly archive old stuff, invalidate wrong stuff, link new connections, your memory becomes unusable within months. Like a cluttered desk, but worse, because the AI cannot tell the clutter from the signal.

Recipe 4.10 covers the actual hygiene routine — the weekly + monthly + quarterly tasks that keep memory healthy long-term. If you are running a memory server and have not done that yet, that is your next step.

The honest take

Memory helps when you need it. It hurts when you do not maintain it. The difference between the two is not the vendor you picked — it is whether you treat memory like a database (something you actively curate) or like a cache (something you set and forget).

Set-and-forget memory will fail you in six months. Curated memory will compound for years.