AI has only increased its use cases over the past two years, getting applied in industries like healthcare, finance, and creative tools. One of the more promising frontiers lies in interactive storytelling, where it can power characters that remember, reflect, and act on their own: forming relationships, developing personalities, and coming up with decisions without scripted prompts.
Cris Lenta saw that potential and built
The Theory Behind LifeSim
When papers like “
But the experiment was done under specific constraints: 25 agents running over two simulated days in a controlled sandbox, with compute costs alone running into thousands of dollars for that limited scope.
What that research did for people like
Every critical subsystem, like memory retrieval, reflection synthesis, and real-time coherence, introduces compounding engineering problems once those numbers grow by orders of magnitude. Latency climbs, token costs multiply, management across asynchronous sessions breaks down, and temporal consistency can become a serious architectural challenge.
But Cris saw this gap between the technical and the practical and treated it as the core engineering project to solve.
His academic background gave him the knowledge he needed. Studying at Ludwig Maximilian University of Munich, he co-authored two peer-reviewed papers published in MIT Press's Artificial Life journal and IEEE Xplore, both focused on self-replicating neural networks. "Prior academic work demonstrated this technology with 25 to 50 agents in controlled research environments," he recalls. "And I sought to turn it into a consumer product where players created tens of thousands of characters."
Memory That Behaves Like Memory
At the center of LifeSim's architecture sits a memory system built on salience-weighted retrieval with temporal decay. Within this system, high-intensity memories persist while routine interactions fade. Each memory stored by a character is scored on two axes: emotional intensity at the moment of formation and recency relative to simulated time. High-salience memories resist decay; mundane interactions (like routine greetings and trivial exchanges) are progressively deprioritized.
The practical effect is that a character remembers a player's first meaningful conversation months of simulated time later but forgets routine greetings from last week. This produces the kind of selective recall that makes relationships feel authentic, directly addressing a typical limitation in naive RAG-based NPC systems, where characters either remember everything with equal weight (which can create an uncanny effect), or forget contextually important interactions altogether, breaking immersion.
The memory system also feeds into autonomous relationship formation. NPCs run independent planning cycles with theory-of-mind modeling, forming opinions and developing conflicts without any player involvement, with relationship trajectories generated using compatibility scoring across personality vectors and shared experience logs. The result is, as Cris explains, “Players discover that characters have history with each other (friendships formed, grudges held), creating a social fabric that exists independent of player action.”
Solving the Time Problem
Another particularly difficult design challenge Cris tackled was temporal coherence: meaning, how agents remain believable across different timescales when simulation speed is fixed. His solution was a system he calls attention-based temporal mechanics, where simulation granularity scales with player attention. Time passes quickly during routine, regular moments and slows during meaningful interactions that will drive the plot.
The system monitors interaction density, emotional valence, and narrative tension in real time. During high-signal moments, the tick rate decreases, expanding subjective time, whereas during low-signal periods, time compresses, and hours can pass in seconds.
Underneath that sits a dual-tick architecture separating player-facing simulation from background world evolution. When a player is not actively observing, NPCs continue pursuing goals on a separate, lower-cost tick cycle. The system maintains serialized world-state snapshots capturing entity positions, relationships, and in-progress NPC goals. When a player goes back into the world, the simulation reconstructs the story from the last snapshot.
This two-layer approach is what makes persistent living worlds feasible without prohibitive compute costs. Background ticks are cheaper to run because of their lower granularity and because they skip rendering-related overhead.
An AI-Driven Living World
The numbers behind LifeSim reflect sustained engagement rather than one-time curiosity. Players have created more than 60,000 characters on the platform, each generating its own memory graph and relationship network. The system manages all of this across asynchronous sessions without manual cleanup. Peak user sessions have reached seven hours, a figure that suggests a lasting level of immersion that doesn’t come from mere chatbot-style interaction.
Building the platform from scratch as a single founder, Cris has already raised $350,000 from gaming-experienced investors, including San Francisco–based Founders Inc, whose leadership includes exits to Twitch, and Paul Braghiel via SMOK VC, an early investor in Unity and Roblox. Cris is now preparing a seed round to scale the team and infrastructure, with plans to open a creator program enabling third parties to build and monetize AI-driven narrative experiences on the platform.
Generative agents are only improving their internal capabilities and the industries in which they can be used, and Cris Lenta's work with LifeSim proves they can survive the leap from research paper to consumer product. What remains to be seen is how far this proposal can grow from here.
This story was distributed as a release by Jon Stojan under HackerNoon’s Business Blogging Program.