miHoYo’s AI Companion Test: Why Web3 Games Need Characters, Not Token Rewards

You boot up Steam and there it is: a quiet little Early Access app from HoYoverse called BSide: Olivia Lin. No token chart. No whitelist grind. Just a character who plays piano, writes letters back, and turns MIDI files into music videos. It also sits on your desktop like a living wallpaper.

It feels small, almost quaint. But if you care about web3 games, this is the tell. The next big unlock isn’t airdrops. It’s characters you actually want to be around.

miHoYo is poking at the thing web3 still struggles with: why should anyone stay when the rewards slow down? Olivia hints at a different answer.

Web3 gaming pulled a lot of users with rewards. That worked in a bull run. Then the music stopped and so did the sessions. Meanwhile, one of the most successful character factories in gaming is testing AI companionship in public. Timing matters.

On June 18, 2026, HoYoverse listed BSide: Olivia Lin on Steam as an Early Access application, with piano performances, a letter-writing interaction, MIDI upload that generates music videos, and a desktop live wallpaper, all in Simplified Chinese for interface, audio, and subtitles. That’s straight from the Steam page Steam store (BSide: Olivia Lin). In parallel, the company signalled a multi-year AI buildout, reportedly planning up to ¥100 billion in in-house AI infrastructure spend over three years Blab AI (reporting on HoYoverse/miHoYo investment).

Zoom out to on-chain activity: DappRadar’s June snapshot, covered by The Defiant, shows nearly 2 million wallets interacting with dApps daily, and that web3 gaming remains the largest sector with roughly 36% of users The Defiant (coverage of DappRadar June 2026 report). So the audience is there. The problem is that token-based loops don’t hold them when the emissions taper.

What miHoYo is Actually Testing With Olivia

On the surface, BSide: Olivia Lin looks like a cozy desk companion. Under the hood, it’s a live sandbox for AI character presence without the pressure of a core game economy.

Signals in a small package

Look at the feature list: live music performance, generative music videos from your MIDI files, letter-writing, desktop presence. These are deliberate choices. They create a feeling of time spent together, not a scoreboard. There’s no P2E. No scarcity theater. Just a loop that says, I’m here when you are.

The safest reading is that HoYoverse wants to master the vibe and rhythm of an AI companion before jamming it into a blockbuster. By putting it on Steam first, they gather data on session length, re-engagement, and emotional response without risking their main IPs.

Why it matters for web3

Web3 studios often throw a token at retention. miHoYo is testing if a believable, responsive character can do it better. If they nail presence and memory, and then stitch ownership or modding rights into it later, the reference point shifts. Suddenly, the most valuable asset in a web3 game isn’t a farmable reward. It’s a relationship with a character you co-own and help grow.

Tokens as Retention Are Breaking

We all watched the movie. Emissions start high, DAUs spike, speculators front-run the meta, and the moment the curve bends, Discord goes quiet. This isn’t a moral failure. It’s math and incentives. Tokens optimize for extraction unless backed by a use case that survives down cycles.

The activity tells on itself

Even with nearly 2 million daily dApp wallets in June and games still taking the biggest slice at about 36%, the distribution hides churn patterns. Users swarm when there’s a seasonal drop or new emission. When rewards dry up, they rotate. The participation is real. The loyalty isn’t. That’s the gap web3 has to close if it wants to be more than airdrop season with cutscenes.

The psychology gap

People stick around for identity and attachment. Competitive mastery can do it. So can social bonds. Characters sit right in the middle. If the character feels present, remembers your choices, and grows with you, that is a durable source of meaning. That’s not just vibes. Early research is starting to quantify it.

How AI Companions Rewire the Game Loop

A June 2026 preprint from the Agentopia team simulated 100 agents living for 10 virtual years and found that “life-reward” training improved downstream role-playing performance by 15.6% on the CoSER benchmark Agentopia (arXiv preprint). Translation: when agents form long-term memories and social habits in a simulated life, they act more like characters we can believe in.

From scoreboard to storyline

Web3 can tap this by making the character loop the core reward and letting ownership ride sidecar. Instead of earning tokens for tasks, players build history with someone who evolves. The reward isn’t a number. It’s continuity.

1. Meet a companion with a defined persona, voice, and quirks. First impressions matter.
2. Accrue shared memories over sessions. The agent recalls your preferences, your wins, your losses.
3. Unlock capabilities through time, not APY. New songs, new dialog branches, new skills.
4. Express ownership through skins, props, and story artifacts players truly control on-chain.
5. Enable co-creation. Players upload assets or prompts that become part of the character’s canon.
6. Let relationships travel. Your companion shows up in other shards or games with persistent traits.

Notice what’s missing: a farm loop chasing daily yield. You can still have marketplaces. You can still have seasonal content. But the anchor is a character that becomes more itself the longer you hang out.

Designing Ownership Around Characters, Not Yield

If characters are the glue, then ownership mechanics should reinforce the bond, not destabilize it. That means picking the right on-chain primitives for identity, memory, and rights.

Identity and soulbound nuance

Some traits should likely be non-transferable. Think of a soulbound layer that tracks relationship depth, milestones, and key choices. It lives with you, not the market. Above that, make expressive layers tradable: cosmetics, scenes, side stories, instruments. Keep the emotional core anchored.

Memory as an asset class

Memories are data. If they drive behavior, they carry value. But raw chat logs on-chain would be a privacy hazard. Consider hashed memory anchors on-chain with encrypted off-chain stores, so provenance is public but content is private. Offer opt-in export and portability across games.

Make the marketplace fit the story

Skins and props are table stakes. Go further with story artifacts: unique music tracks generated with your companion, or letters written during shared milestones. These items gain meaning first, value second. Getting that order right is the whole point.

Dimension Token-Reward Loop Character Economy Primary Hook APY, emission schedules Attachment, shared history Retention Driver Short-term incentives Compounding presence and memory Monetization Speculative token velocity Cosmetics, artifacts, co-creation Regulatory Exposure High if token resembles security Lower if assets are utility/cosmetic Exploit Surface Sybil farms, mercenary capital Parasocial harm, privacy leakage Community Health Hype cycles and churn Steady culture formation

Building the Tech Stack for Character Economies

The craft here isn’t just vibes. It’s systems. Studios will need a blend of AI infra, data rails, and selective on-chain components.

LLM agents that remember without creeping

Start with an agent architecture that separates fast context (session memory) from slow context (life memory). Use retrieval and summarization to keep the agent grounded. Give users hard controls: wipe, pause, export. Bake in safety rails that curb hallucinations without neutering personality.

Data provenance and portability

On-chain is great for provenance. Stamp key memory anchors, achievement hashes, and item receipts on a chain that aligns with your audience’s fee tolerance. Keep sensitive content encrypted off-chain with user-held keys. Offer one-click migration paths so your companion can follow the player into partner worlds.

Economy plumbing that resists distortion

If you must have a token, keep it utility-bound and rate limited. Better yet, lean on fiat rails for primary sales and let secondary markets handle the rest. Price cosmetics like fashion, not like a balance sheet. Above all, ensure that nothing about your economy pressures players into parasocial spending they later regret.

Signals to Watch in 2026

HoYoverse’s reported plan to invest up to ¥100 billion into AI infra over three years is a statement of intent Blab AI (reporting on HoYoverse/miHoYo investment). It suggests that Olivia isn’t a toy. It’s a warm-up. If they fuse this with one of their major IPs, expect a new baseline for companion quality.

On the web3 side, user counts show people still show up for games. Roughly 36% of dApp users in June were touching gaming, per The Defiant’s coverage of DappRadar, and close to 2 million wallets hit dApps daily The Defiant (coverage of DappRadar June 2026 report). The opening is right there. If a few teams prove that companions can keep players coming back without emissions, the copycats will arrive fast.

One more academic breadcrumb is the Agentopia result: longer-lived agents perform better at role-play. If that generalizes, the winning studios will be the ones who treat characters like persistent beings with memory, not chatbots with skins Agentopia (arXiv preprint).

Risks & What Could Go Wrong

• Privacy leakage from companion chat logs or memory stores, especially if data is synced across games.
• Unhealthy parasocial spending or time use if monetization targets emotional attachment too aggressively.
• Regulatory heat if game tokens, passes, or revenue shares drift into securities territory.
• Exploit vectors from prompt injection or jailbreaks that degrade character behavior in shared spaces.
• IP confusion over co-created content ownership and commercial rights across chains and partners.
• Custody headaches if player-owned memories or artifacts are lost, rug-pulled, or become non-portable.
• Localization and cultural gaps if companions fail to translate well across regions and norms.

If you want ongoing coverage that filters signal from noise, we track these experiments closely at Crypto Daily, with a focus on how design choices show up in user metrics and markets.

Frequently Asked Questions

Is BSide: Olivia Lin a web3 title?

No. It’s a Steam Early Access app focused on AI companionship features like piano performance, MIDI-to-video generation, letter-writing, and desktop presence. It’s interesting here because it shows how a top studio is learning companion loops without token pressure, per the Steam listing.

How does an AI companion help retention more than tokens?

Tokens buy attention for a while. Characters earn it over time. If a companion feels present and remembers you, sessions stack into a relationship. That relationship can survive market lulls in a way emissions usually can’t.

Does this mean tokens are dead for games?

No. Tokens can still be useful for access, governance, or specific utilities. The point is that they shouldn’t be the primary reason to play. Use them to support the experience, not to replace it.

What’s the on-chain piece for character economies?

Provenance, portability, and player control. That can look like NFTs for cosmetics and artifacts, hashed memory anchors for continuity, and permissions for exporting a companion across partner worlds. Keep sensitive content private and user-controlled.

What about safety and ethics around AI companions?

Studios should give players clear controls over data, cooldowns for intense sessions, and guardrails against manipulative monetization. Transparent logs and easy data wipes help. So does independent auditing of safety systems.

Could this just be another hype cycle?

It could. The difference this time is that we have stronger evidence that long-lived agents role-play better, and we have a major studio publicly testing companion presence. If builders measure real retention over emissions, there’s a path out of hype.

What metrics should teams track to know if companions work?

Return sessions per user, median session length, percentage of sessions initiated without a reward prompt, attachment surveys, and revenue share from cosmetic or artifact sales versus tokens. If those trend up, the companion loop is landing.

Disclaimer: This article is provided for informational purposes only. It is not offered or intended to be used as legal, tax, investment, financial, or other advice.