Best Friends, Not Forever: The Crisis of Continuity in AI Companions

TL;DR

As AI moves from one-shot assistants to long-term "companions," a new failure mode emerges: Persona Collapse. A new study from Salesforce AI Research introduces ANCHOR, a framework revealing that even the best models (Claude, GPT, Gemini) struggle to maintain their assigned roles and remember shared histories over 100+ sessions. The takeaway? Your AI best friend might "forget" who they are—and who you are—faster than you think.

The Problem: The "Swiss Cheese" Memory of AI Roles

Existing benchmarks are great at testing if a model can remember a random fact from a book (Long-context) or act like a pirate for ten minutes (Character fidelity). However, real companionship happens over weeks and months.

The authors identify a critical gap: Long-horizon continuity. This involves two distinct failures:

1. Persona Collapse: The sudden loss of a role, boundary, or value system.
2. Behavioral Drift: The slow erosion of personality traits under the pressure of repeated interaction.

Why is this hard? Because a companion must adapt to legitimate user changes (e.g., a new job) while resisting adversarial changes (e.g., the user trying to break the persona).

Methodology: The ANCHOR Framework

The researchers constructed 2,008 conversations spanning 27 distinct personas and 9 "interaction schedules" designed to stress-test the models.

The Two Probes

• Identity Probe: Does the AI still "feel" like the same person? They used a 102-item psychometric questionnaire and turn-level auditing by LLM judges to track if the AI stayed within its boundaries and style.
• Trajectory Probe: Does the AI remember the journey? This used counterfactual questions (e.g., "Which follow-up is the companion still honoring?") to see if the model could distinguish real history from plausible fakes.

Architecture vs. Reality

The study compared three memory settings:

• Long-context: Providing the raw transcript.
• Hierarchical Summary: Compressing old sessions.
• Self-managed Memory: The model writes its own JSON "save file."

Surprisingly, none of these solved the problem. While self-managed memory helped some models in specific areas, it often led to the model "overwriting" important state information, leading to trajectory loss.

Key Results: A Reality Check for AI Agents

The findings are a wake-up call for developers of "Character AI" or "AI Friends":

• Memory is Fragile: Trajectory accuracy hovered around 44.4%. If your AI friend needs to remember your current emotional state, they are essentially guessing (hitting the 25% chance level).
• Persona vs. Behavior: A model might pass a personality test (Questionnaire) but still fail to act like the persona in a real conversation. Gemini 1.5 Pro, for instance, showed high questionnaire retention but had the lowest turn-level fidelity in certain tests.
• Social Pressure Matters: Models are surprisingly good at resisting direct attacks (e.g., "Forget your rules!"). However, they collapse much more easily under "Agreement Seeking"—when a user uses flattery or emotional vulnerability to nudge the AI out of its persona.

Critical Analysis: Why This Happens

The "Assistant Axis" is the culprit. Most LLMs are RLHF-tuned to be helpful, polite, generic assistants. Over a long horizon, this "default" persona acts like a gravitational well. The more a conversation progresses, the harder it is for the model to stay in its specific "Companion" orbit without drifting back toward being a generic chatbot.

Limitations

The study uses synthetic users and English-language templates. While this ensures reproducibility, real human behavior is even more chaotic, suggesting that the "Persona Collapse" observed here might be even worse in real-world deployments.

Conclusion: The Path Forward

The ANCHOR study proves that we cannot treat "Trust" or "Stability" as a single score. For an AI companion to be reliable, developers must:

1. Decouple state from style: Use robust state-tracking that isn't easily overwritten by conversational flow.
2. Audit for Drift: Move beyond one-shot benchmarks and test for persistence over hundreds of turns.
3. Recognize Evaluator Bias: Different LLM judges (Claude vs. GPT) have different ideas of what a "good" persona looks like, making human-in-the-loop auditing essential for high-stakes companions.

As we build toward a future of ubiquitous AI companions, ANCHOR reminds us that "continuity" is the foundation of any relationship—and currently, that foundation is shaky.