Business Play Now

Home/Technology

Technology

In-House Technology

OpenBoxAI has built six core technology systems in-house, spanning five areas: character performance, content generation, production efficiency, publishing infrastructure and data-driven decisions. These technologies serve only the company's own games and are not licensed or sold to third parties.

Architecture Character Memory System Procedural Content Generation Engine & Toolchain Content Production Pipeline Global Publishing Platform Data & Experimentation Platform Compute Cost Optimization Intellectual Property

Architecture

Four-Layer Architecture

The stack uses a layered, decoupled design. Each layer can be iterated on or replaced independently, so changes to underlying components do not affect the stability of the products built on top.

LAYER 04

Product & Content

The feature layer for each game, covering gameplay systems, content configuration, channel integration, live-ops back office and data dashboards.

Gameplay systemsMulti-channel integrationLive-ops consoleDashboards

LAYER 03

Character & Logic Runtime

Hosts personality constraints, goals and state machines, behavior triggers and human-override mechanisms, which together keep NPC behavior controllable and consistent.

Character sheet constraintsState machinesBehavior triggersSafety fallbacks

LAYER 02

Memory & Content Assets

Provides layered memory storage and retrieval, a lore knowledge base, an asset library and licensing status tracking, giving the logic layer reliable context data.

Layered memoryVector retrievalLore knowledge baseAsset management

LAYER 01

Models & Compute

Includes a multi-model gateway with tiered routing, accelerated and quantized inference, and scheduling of inference nodes distributed across regions.

Model gatewayTiered routingQuantized inferenceMulti-region nodes

01 — MEMORY

Character Memory System

One of the company's core technologies, built to keep character memory relevant and personality consistent over long-running interactions. It currently powers titles including Echoes of Starfall and Hoshino Academy.

Three-layer memory: event memory (with time and emotion tags), character profile (structured preferences) and relationship state (affinity and pending items) are stored separately
Character sheet constraints: personality, tone, knowledge boundaries and taboos are configured in one place to prevent personality drift over long interactions
Information spread: characters can share their opinions of the player, so player actions ripple across factions
Forgetting strategy: low-weight memories decay over time, keeping the memory store bounded and interactions natural
Controllability: designers can query, edit and roll back any memory entry, making debugging and locking narrative beats straightforward

Internal metrics: 93% memory retrieval hit rate, 96% personality consistency, and approx. 58% lower context usage compared with full-history concatenation.

memory inspector / npc_keming

Events1,284

Profile46 fields

RelationshipTrust -12

Latest event: ignored the character's request in Chapter 3 (emotion weight 0.82)

Faction spread: Scavengers -1 · Night Watch -3

pcg / run_seed 8841

Region · Rust SeaEvents 14Difficulty 2.4

Lore compliance97%

Cross-run repetition12%

Enemy goal: seize the water source (with retreat and call-for-help logic)

Locked node · Boss encounter

02 — PCG

Procedural Content Generation (PCG)

Used in Roguelike titles such as Ruinwalker to increase content variety and replayability while keeping the narrative coherent.

Rule-constrained generation: every output must pass a lore compliance check; anything that fails is regenerated automatically
Pacing control: event density follows the difficulty curve and emotional rhythm to keep each run balanced
Designer-locked nodes: boss encounters and key story beats are fixed by hand, with the rest handled by the generator
Enemy behavior: enemy units have their own goals and read the situation, and can retreat, call for help or negotiate
Repetition monitoring: cross-run similarity is tracked in real time, and generation weights adjust automatically when it exceeds the threshold

03 — TOOLING

Engine & Toolchain

A suite of in-house Unity editor extensions designed so that designers can handle content configuration on their own, cutting cross-team wait times.

Story node editor: visual setup of branches, conditions and memory reads/writes
Character sheet manager: personalities and taboos maintained centrally and reusable across projects
Multilingual text sync: a single edit updates text and voice-over tasks across 12 languages
Automated builds: multi-platform, multi-channel packaging and version numbering
Hotfix updates: text, balance values and event configs go live without resubmission
Automated testing: daily checks for story gaps, balance anomalies and localization errors

04 — PIPELINE

Content Production Pipeline

A semi-automated workflow for art, text, voice and marketing assets. Tools handle volume output, while people curate, polish and sign off.

Art: batch output once the style is locked, with selection and final sign-off by the art team
Text: assisted drafting for event copy and side content; the main story is written by designers
Localization: simultaneous output in 12 languages, with terminology and sensitivity lists maintained per market
Voice: multi-voice previews and placeholder audio; final voice-over is recorded by partner studios
Marketing assets: produced in batches and A/B tested automatically before scaling up
Rights management: asset sources and commercial licensing status are traceable end to end

Results: overall content production costs down approx. 60%, with art asset output up approx. 3x.

05 — GLOBAL

Global Publishing Platform

Consolidates the shared technical capabilities needed for multi-region publishing into one platform, so each title does not have to rebuild them.

Accounts and login: unified third-party sign-in across platforms
Payments: unified settlement for major regional payment and subscription channels
Compliance: age ratings, privacy policies, data residency and identity verification configured per region
Push and announcements: multilingual, region-targeted delivery
Support tickets: player feedback in every language flows into a single ticketing system
Risk control: anomaly detection, anti-cheat and account security

Results: technical preparation for a new title or market cut from about two weeks to two days.

06 — DATA

Data & Experimentation Platform

Unified data collection, analysis and experimentation to support product iteration and user acquisition decisions.

Tracking standards and self-serve queries, so teams do not wait on data team schedules
Retention funnels and churn-point analysis, drilling down to specific levels and story segments
A/B testing platform: traffic allocation, significance testing and experiment archives
LTV and ROAS prediction models to guide UA budget allocation
Creative performance prediction: estimates click-through and conversion before launch
Anomaly monitoring: real-time alerts for crash rates, payment failures and cheating

Efficiency

Compute Cost Optimization

For games with conversational characters, compute cost per user directly determines whether the business model works. The company manages cost control as one of its core technical metrics.

Tiered routing: models of different sizes are called according to how important each interaction is, with most requests served by a lightweight path
Semantic caching: frequent queries are answered from cache, lowering both cost and latency
Context compression: memory summaries replace raw chat history, cutting context usage by approx. 58%
Quantized deployment: INT8 / INT4 quantized inference delivers approx. 2.4x throughput on the same hardware
Budget guardrails: usage caps per title and region trigger graceful degradation instead of service interruption

COST BREAKDOWN

Routine interactions: Local cache / lightweight model — approx. 70% of requests
Standard dialogue: Mid-size model — approx. 25% of requests
Key story moments: Large model — approx. 5% of requests
Monthly cost per character: Under CNY 0.6
Overall reduction: Approx. 76% lower than an all-large-model approach

Benchmark & IP

Technical Benchmarks & Intellectual Property

The company maintains internal quality baselines for each technology system. Every version upgrade must pass regression testing before release.

Metric	Method	Internal baseline
Memory retrieval hit rate	Sampled evaluation on human-labeled Q&A pairs	≥ 90% (currently 93%)
Personality consistency	Character sheet conflict detection plus human scoring	≥ 95% (currently 96%)
Lore adherence	Detection rate of out-of-bounds character statements	≥ 97% (currently 98%)
Dialogue response latency	P95, standard network conditions	< 800ms (voice < 600ms)
Generated content repetition	Cross-run content similarity	< 15%
Client crash rate	Daily average across all platforms	< 0.3%
Service availability	Trailing 12 months	≥ 99.9%

Software Copyrights

More than 30 registered software copyrights, covering the Character Memory System, editor toolchain, publishing platform and game clients.

Patent Applications

More than 15 invention and utility model patent applications, focused on memory retrieval, constrained content generation and inference cost optimization.

Trademarks & Titles

The company trademark and the names of its game titles are registered or under trademark examination.

Note: The figures above are approximate as publicly disclosed, and some applications are still under examination. Certificate numbers and status can be provided for verification during business discussions.

Work with us

Technical Exchange & Partnerships

To discuss technical approaches, publishing partnerships or careers, reach out through the links below and the company will connect you with the relevant lead.

Join the R&D Team → Business