Klang Details the Gemma-Based AI Behind SEED's Persistent Characters
Klang Games CEO Mundi Vondi told 80 Level on August 5 that SEED runs its persistent characters with a modified Gemma-based model on Klang's own hardware. Each Seedling maintains an experience-built knowledge graph, while controls such as a social battery limit interaction pressure. The design keeps player-owned characters working and socializing when their players are offline.
Klang Games has disclosed more of the AI architecture behind SEED, its persistent society simulator. In an interview published by 80 Level on August 5, CEO Mundi Vondi said the studio uses a modified model based on Google's open-source Gemma family, runs inference on its own hardware, and stores an evolving knowledge graph for every character.
SEED's characters, called Seedlings, belong to players but remain active when those players log off. They work, spend money, maintain relationships, use services, and contribute to player-built cities. That persistence is central to the game's design: the simulation continues instead of pausing at the end of a player session.
Memory is part of the simulation
Vondi described each Seedling's knowledge graph as subjective rather than global. A character's graph grows from what it sees, hears, and experiences, so two characters can hold different beliefs about the same relationship or event. That asymmetry lets the simulation produce misunderstandings and discoveries without giving every character access to the same world state.
The interview also adds details absent from Klang's earlier public overview. Klang says it modified a Gemma-based model, built a custom inference pipeline, and operates it on company-controlled hardware. The studio stores expanding character knowledge without keeping the entire graph in memory and uses controls including a "social battery" to reduce the risk of players overwhelming characters with continuous interaction.
Klang's May 2025 post had already described generative AI as one layer of its broader Game AI system, including free-form chat, evolving personalities, and player-influenced stories. The August interview connects those product claims to a more specific deployment pattern: model inference is coupled to durable simulation state rather than used as a standalone dialogue service.
The production test is coherence
The architecture does not guarantee convincing behavior. PC Gamer's July 24 hands-on account praised some question-answering interactions but criticized repetitive, immersion-breaking dialogue. That is anecdotal product coverage, not a systematic evaluation, yet it points to the central engineering risk: fluent text can still conflict with character identity, world state, or player expectations.
For teams building persistent agents, SEED illustrates why memory, state validation, inference cost, and interaction limits matter alongside model quality. Klang has not published latency, throughput, cost, safety-evaluation, or long-horizon consistency metrics, so the system's reliability at the claimed scale remains unverified publicly.
Key Points
- 1Klang says SEED uses a modified Gemma-based model, a custom inference pipeline, and company-operated hardware for persistent character behavior.
- 2Each player-owned Seedling builds a subjective knowledge graph from experience, allowing beliefs and relationships to diverge across characters.
- 3The public evidence describes memory and interaction controls but does not disclose latency, cost, safety, or long-horizon consistency benchmarks.
Scoring Rationale
The interview provides unusually concrete deployment details for a live persistent-agent game, including model family, hosting approach, knowledge graphs, and interaction controls. Public evidence still lacks scale, cost, latency, safety, and reliability benchmarks.
Sources
Public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
