Modal: A State-of-the-Art Reward-Driven Memory Towards Personal Intelligence

Today, we're excited to share a significant breakthrough in AI memory systems using Modal: our reward-driven memory system. Our team at Elicit Labs has achieved 83.6% accuracy on the LoCoMo benchmark with GPT-4.1-mini as answering model, matching in-context learning performance while using 15x less context on an only 29k token size conversation. This represents an advancement in making AI personalization possible at scale.

What makes this particularly significant is that our system continuously updating its understanding from every interaction. It extracts salient features that define a user's style, intent, and preferences, and assigns rewards accordingly.

Elicit Labs: Building Personal Intelligence

At Elicit Labs, our mission is to fundamentally elevate Human–AI interaction. We believe that intelligence is not just about producing correct answers, it's about understanding people. A truly intelligent system should be aware of who it is interacting with, recall shared history, infer preferences and values, and adapt its behavior accordingly.

We also believe that AI systems should maintain evolving behavior grounded in memory, preferences, and learned patterns of interaction rather than static prompt templates. An AI that can remember, refine, and respond consistently over time isn't just more useful, it's more human-compatible.

Our approach centers on building user-conditioned models that shape AI behavior based on individual user preferences. The challenge: training such models requires not only novel architectures but also large and diverse sets of user-specific preference data, but explicit preference labeling is expensive, slow, and cognitively unnatural.

This is where Modal becomes essential. Our research demonstrates that by building stable, selective memory representations of user interactions, we can derive implicit preference signals directly from natural conversation, no explicit feedback required. This memory foundation enables us to train reward models that drive truly personalized AI systems.

Our goal is clear: build the future of Human-AI interaction for all AI systems that service humans, from agents, to healthcare, robotics, and beyond.

Modal: Our Approach

Our breakthrough builds on two core innovations: a biologically inspired memory partition system and a heavy reliance on reward models to guide memory formation and retrieval. Drawing from neuroscience research on how human memory organizes and consolidates information, we developed a partition structure that naturally separates different types of learned information.

The reward model framework is central to our approach. Rather than treating all memories equally, our system uses learned reward signals to determine what information is valuable to retain, how it should be stored, and when it should be retrieved. This allows the system to prioritize memories that drive better personalization outcomes while efficiently discarding noise.

Humans forget, as should AI systems. Reward-weighted temporal forgetting is a key component of our approach, dynamically pruning memories that are no longer useful to the user while reinforcing the most relevant ones while maintaining a contextualized exploration strategy.

Crucially, this architecture learns continuously from every interaction while remaining computationally efficient enough to serve millions of concurrent users.

Evaluation: The LoCoMo Long-term Memory Benchmark

We evaluate Modal on Locomo, a benchmark designed to test whether a model can retain and apply user-specific context across long-term conversations. Unlike short-context dialogue tasks, Locomo requires the model to:

• Track personal events, beliefs, preferences, and relationships across sessions
• Resolve references to past experiences and emotional states
• Reject incorrect or misleading inferences when evidence is insufficient

Locomo frames this as a question-answering task across five reasoning categories:

• Single-hop: Answerable from one conversation turn.
• Multi-hop: Requires synthesizing information across multiple sessions.
• Temporal reasoning: Requires understanding timelines and evolving circumstances.
• Open-domain knowledge: Requires combining conversation details with world knowledge.

Results and Impact

Our system achieved 83.6% accuracy with GPT-4.1-mini, a 2% increase on the in-context learning baseline while using just 2.1k tokens compared to ICL's 29k tokens, a 14× reduction in context requirements with no loss in performance. With GPT-4o-mini, we achieved 80.7% accuracy, substantially outperforming specialized memory systems like Zep (76.5%) while maintaining comparable context efficiency.

Organizations will be able to deploy personalized AI at scale without the computational costs that have historically made such systems prohibitive. Memory updates happen in real-time, personalization improves continuously, and the entire system remains efficient enough for production deployment with millions of users.

From Memory to Personalized Rewards

Our memory system isn't just about remembering information, it's the foundation for learning user preferences and embeddings at scale from conversational traces.

Once memory stabilizes, we extract implicit pairwise preferences from conversations by generating counterfactual candidates with high confidence, leveraging memory to create preference pairs. This enables training user-conditioned reward models that guide generation during post-training or at inference time, achieving personalization that adapts to individuals, not population averages.

What's Next

We're launching a series of additional results and releasing an SDK that will make our memory system accessible to the developer community. By putting these tools in the hands of builders, we aim to accelerate innovation in personalized AI applications and discover new use cases across domains from healthcare to robotics.

We're actively seeking collaborators focused on personalization evaluation and domain-specific alignment, particularly in settings where capturing real user intent is critical. We'll be publishing detailed technical papers in the coming weeks and sharing opportunities to build with state-of-the-art AI memory.