Skip to main content
Back

LifeSide Exposes How Badly AI Companions Fail at Long-Term Memory

LifeSide exposes a critical gap: frontier LLMs that ace memory benchmarks still fail at sustained emotional understanding across sessions

Paper, Full Access
  • LifeSide is a new benchmark for evaluating AI agents as lifelong digital companions, covering memory, emotion, and privacy across multi-session interactions.
  • Built on 2,000 personas and 111K tasks spanning simulated 24-to-36-month user timelines with evolving life events and emotional states.
  • Frontier models score only 41.24% on episodic recall and below 37% on emotional companionship -- exposing major gaps hidden by existing benchmarks.
  • A surprising finding: RAG (retrieval-augmented generation) degrades empathy by up to 22% -- more memory access hurts emotional responsiveness.
  • The benchmark introduces a Memory-Emotion-Environment (MEE) loop that tests agents under partial observability, where user internal states are never directly revealed.
  • First benchmark to systematically evaluate privacy boundary tracking -- whether agents respect shifting user preferences about sensitive information over time.

AI companions are everywhere now -- Replika, Character.AI, Pi, and a growing list of products promise to remember you, understand you, and be there for you over time. A year ago, users asked whether an AI companion felt natural. The sharper question now is whether it remembers you. But a new benchmark called LifeSide suggests that even the best frontier models are nowhere close to meeting that bar -- and that our current evaluation tools have been hiding this failure all along.

The benchmark gap nobody was measuring

The problem is that existing evaluations test memory and emotion as two completely separate things. Memory benchmarks primarily test factual retention, while emotional support benchmarks isolate empathy to short, self-contained dialogues. That's fine for a chatbot that answers questions, but it misses everything that makes a companion actually useful over months or years.

For example, MADial-Bench integrates memory into dialogue but lacks environment dynamics; ES-MemEval explores personalized long-term support but assumes full observability of the user's state. Existing benchmarks fail to evaluate the core challenge of real-world companionship: maintaining a unified, evolving understanding of a user amidst changing external conditions and incomplete information.

Real companionship is messier. A user's mood on a Tuesday is shaped by what happened the previous Friday, what they didn't say last month, and whether they're going through a stressful period at work. A true companion agent must go beyond task execution or one-off question answering; it must provide sustained, deeply personalized emotional and cognitive support over months or years. No existing benchmark was testing for that -- until now.

Evolution from single-episode emotional support to memory-enabled support to long-term emotional companionship

What LifeSide actually tests

LifeSide is a benchmark centered on multi-session Memory-Emotion-Environment (MEE) loops. By modeling users as persistent worlds with layered profiles and event trajectories, it uses multi-agent simulation to project environmental dynamics into dialogue, preserving the critical gap between latent thoughts and observable expressions.