I grew up on Naruto, Slam Dunk, and Bleach. Anime shaped my childhood.

But when a crypto startup leans hard into anime branding, my instinct is to ask: is this a serious team, or are they just cosplay with a ticker?

Rei Network kept showing up anyway. Enough that I had to look closer.

In November 2024, a new handle appeared on crypto Twitter: @0xreisearch. The bio read simply, “pushing boundaries,” with a link to reisearch.box. No photos. No token price threads. Just posts about memory graphs, inference latency, and something called the Bowtie Architecture.

“Reisearch” keeps his profile minimal. He writes about quantum physics, Bayesian deep learning, and EVM internals. Friends recall an early career in statistical physics, followed by a pivot into Solidity security audits.

On Telegram, he filled us in with more detail: a dual background in physics and chemistry, a decade in AI, including founding and scaling startups, an advisory role with a legal-tech AI company, and four years of direct blockchain experience.

Rei Network started with a modest idea: make AI run deterministically so it could plug cleanly into blockchains. That meant building a toolkit for AI × blockchain. But the ambition grew.

The team realized that intelligent agents could create value both inside and outside crypto. And that LLMs alone weren’t enough. They’re brilliant at text, brittle at reasoning. Wrappers couldn’t fix the gap. To move forward, Rei had to build its own architecture.

Across all of it runs a single obsession of giving AI what it has never had before: persistent memory.

Nine months later, they released Rei Core. And today, 23 people are working in their team to make it a reality.

Rei Network is not pitching another fast EVM sidechain. Lord knows we’ve had enough of those. It wants to be a labs where software agents remember what happened yesterday and rewrite their own code tomorrow. If it succeeds, you could launch a trading agent this week and watch it rebuild its own risk model the next.

What makes it worth watching is the way it is being built: the experiment is being run in public, on-chain, and almost entirely with community funds. No VCs. The vibe feels less like a startup, more like a science lab testing bold hypotheses.

In a cycle flooded with fluffy AI startups, this makes Rei one of the more interesting experiments to watch.

An excerpt from one of their recent interviews:

So what does Rei Network…actually do? Let’s start with AI’s biggest problem.

I keep hearing the same praise for large language models: they sound brilliant, they ship code, they write heartfelt birthday notes. Then the praise collapses when the model hallucinates or gives you a wrong answer. The problem boils down to two habits I have come to expect every time I open a new chat window: Goldfish Memory and Peacock Confidence.

Goldfish Memory shows up the moment the model forgets yesterday’s thread and asks me to restate my goals, or gives me an awkward answer because it doesn’t have context. Peacock Confidence follows close behind, filling the gap with flawless prose that masks missing facts.

A doctor who practiced medicine this way would appear masterful during the consult and still kill the patient. That’s the uncomfortable reality of AI today.

If you’re just cranking out summaries or tweet threads, forgetfulness is annoying. Hand the same AI a research portfolio or an options book, and that gap becomes existential. A trading agent that forgets why it made a bad trade last week will replay it, losses included.

An agent without memory is an autopilot ship with a broken compass. Give it weeks at sea and you end up on the wrong continent.

So, devs try to pack entire histories into ever-larger context windows. Google’s Gemini 2.5 has a 1 million token window, which is enough room to fit every word of the seven-book Harry Potter series into a single prompt. But size alone does not create understanding. The data sits adjacent to intelligence rather than inside it.

As Andrej Karpathy reminds us, “Just because you can feed in all those tokens doesn’t mean that they can also be manipulated effectively by the LLM’s attention and its in-context learning mechanism for problem solving.”

Most “memory” solutions in AI today are tactical patches designed to make LLMs look consistent without giving them any internal continuity.

Let’s take them one by one.

Each method patches a symptom without fixing the disease. They simulate coherence instead of delivering it. Stateless models pretending to think.

Rei steps in precisely here. It treats memory as architecture. The team asks how an agent can remember, reason, and evolve the way a seasoned colleague does, then builds from that point outward.

Everything else in the Rei network flows from solving Goldfish Memory without inviting Peacock Confidence to the party.

In the human brain, vision, memory, and reasoning are each handled by their own region, stitched together through constant coordination. The magic comes from the handoffs.

Rei’s architecture follows the same principle. Specialized modules take on specific tasks like memory, planning, and perception. A central reasoning engine integrates them. Cross-checking reduces hallucinations. Persistent memory keeps agents grounded.

The result: Rei Core, a synthetic cognitive system built to think in modules.

Rei Core has three major subcomponents:

Traditional vector databases work like filing cabinets. They store facts and retrieve them when called. Bowtie behaves differently. It acts more like a neural system that reshapes itself as it learns.

The name comes from its design: a three-part pipeline that forms a bowtie-shaped flow.

Left Wing: Semantic Network = Meaning in Human Terms

This layer stores meaning in human words. It builds connections between explicit concepts, like how “bird” → “flight” and “feathers.”

Center: Core Distillation

Strips away noise while preserving essence. Complex phrases such as “distributed autonomous navigation” can be reduced to their fundamentals without losing the idea.

Right Wing: Feature Network = Meaning in Mathematical Patterns

This side handles patterns, not words. It works in high-dimensional vector space to detect statistical patterns across domains. This is how the system finds non-obvious analogies. Like connecting “viral marketing” to “epidemiology” by recognizing the same underlying dynamics of spread and contagion.

The power comes from how these two wings interact. One encodes meaning, the other encodes mathematical structure. So when the semantic side processes “bird” and the feature side sees a match with “airplane,” the system can generate an emergent, creative concept like “biomimicry in aviation”, even if it was never explicitly trained on that link.

Bowtie also reduces hallucinations. The semantic and feature networks act as cross-checks. If a proposed output doesn’t align across both forms of representation (meaning and structure), it gets filtered out.

Rei’s architecture mirrors an important area of research in AI known as neuro-symbolic integration. The idea is to combine:

The hope is to get the best of both worlds: fast statistical recognition and structured reasoning. Academic surveys as recent as 2024 describe this as an open, grand challenge. In that sense, Bowtie’s dual-channel memory is scientifically plausible, and Rei’s framing echoes what multiple university labs are attempting today.

Bowtie stores memory. The Reasoning Cluster decides what to do with it. It acts as the executive of Rei’s synthetic brain.

The Cluster breaks down queries into logical parts. It assigns tasks to the best module. It links memories through both semantics and structure. It favors smaller models when accuracy is equal, valuing efficiency over raw scale. And it keeps a living concept graph that rearranges as new information arrives.

Before sending an output, the Cluster runs a calibration step. Multiple reasoning paths are checked against each other. If they diverge, the system seeks verification or flags the uncertainty instead of producing confident errors. That feedback loop reduces hallucinations.

Rei avoids the trap of using one giant model for everything. Instead, the Switchboard manages a roster of specialized models, each chosen for a specific task.

The Switchboard takes the high-level plan from the Reasoning Cluster and executes it. It breaks complex queries into smaller steps, routes them to the right model, and coordinates them in parallel. It also manages resources to maximize efficiency.

Currently, the ecosystem includes models like statistical predictors for numerical tasks, perception models for vision and audio, and domain-specific models like Hanabi-1 for financial prediction.

This modular design allows Rei or outside contributors to add new specialists over time without rebuilding the system. The Switchboard can grow into a large catalog of experts, each integrated into the same cognitive loop.

Good news for most of us: All of Rei’s cognitive machinery runs under the hood.

What users see are Units, which are persistent AI agents powered by the Core.

A Unit is a personalized brain-in-a-box that learns in real time. It grows from every conversation, builds a personality shaped by your feedback, recalls past interactions with perfect fidelity, and strengthens its reasoning by weaving new conceptual links over time.

To make Units accessible, Rei built the Reigent Factory. It is a platform where anyone can spin up Units without needing technical depth. Configuration is simple. Built-in tools include market analysis and image generation. Developers get APIs for deeper integration.

A Unit can be a personal research partner for a non-technical user or a component in a complex agent system for a developer. The difference from today’s chatbots is continuity. Traditional AI resets after each session. Units grow with you, learning your preferences, goals, and working style.

Every interaction with a Unit runs through the full stack:

The result is a pipeline that not only responds but also grows sharper with each exchange. Units are more like colleagues who never forget, rather than chatbots.

This architectural design shows how memory evolves like a human brain, through clear stages:

The system tracks strength. Irrelevant facts fade; frequently used knowledge hardens into core belief. Agents learn lasting preferences, adjust behaviour after repeated feedback, and replace outdated assumptions when new evidence prevails.

An agent that can revise its original worldview. Powerful? Yes. Dangerous? Also yes.

Normally, you’d teach an LLM new behavior by fine-tuning it. That means collecting data, running training loops, waiting hours (or days), and hoping the model internalizes what you wanted, without forgetting everything else.

Rei skips that entire cycle.

Instead, agents “train” at inference, meaning they learn in real time by interacting with users and environments. It is conceptual memory evolving with use.

Here’s how it works under the hood:

This is compound intelligence in action, where each success sharpens future reasoning. Concepts self-organize. Inference paths optimize themselves. Patterns emerge even if the user never explicitly states them.

While the technology is impressive on paper, the real test is adoption.

Rei's Factory, their Unit (agent) creation platform, is still invite-only, but even under restricted access, demand has been surging. Early on, usage spiked 3X on Core requests and 5X on Bowtie queries, with image generation capped by GPU allocation limits , which is a good problem to have.

By late July, as the closed beta entered its final stages, activity hit new highs. Peak usage crossed 23,000 daily queries (including API calls), up from just ~5,000 two weeks prior, directly correlating with new invite waves and waitlist activations.

Stored concept logs show these users leaning hard into real-world, high-value tasks:

Users are testing edge cases and pushing the system's boundaries. Rei treats this closed test as a dress rehearsal for open beta. The team collects detailed feedback, tracks usage patterns, and scales infrastructure in step with the load.

We secured an invite to test the platform firsthand and built a Unit. The Unit creation process is surprisingly straightforward. The interface is clean and minimal. A stripped-down interface asks for a name, optional tags, and a behaviour prompt. Power users can open advanced settings, yet the basic setup finishes in minutes.

Once live, the Unit learns from every line of chat. It keeps a perfect record of your preferences, tone, and favourite analogies. Ask it to store an investment thesis; it does. Explain your research method; it adapts. Correct its errors; it updates the underlying concept graph. Over time, the Unit forms a distinct personality and a knowledge base shaped around the way you think.

Each Unit ships with an API key. Developers can wire it into trading engines, workflow automations, or bespoke tools that follow standard function-calling schemas. The Unit builds memories of these external tools just as it does with its native skills, so learning persists across every environment you plug in.

Rei’s development calendar moves fast. In twelve months, the team delivered three major versions of Core and loaded several more into the queue.

The latest update, Core 0.3.3, introduced zero-decay memories. These “Primordials” let Units preserve critical facts and instructions indefinitely. Living directly inside Rei’s hypergraph, they never degrade, always trigger through semantic context, and actively shape reasoning paths. In practice, that means a Unit can recall your compliance rules or technical specs with perfect fidelity, the 100th time as clearly as the first.

In July, they released their Chain Data Engine Beta, a unit-level upgrade that pipes in live feeds from CoinGecko, Nansen, Birdeye, DexScreener, and DeFiLlama. The engine boosts capture accuracy and widens analytical reach, which enables Units to surface richer views of on-chain behavior and market shifts.

Looking ahead, Rei has previewed R00Ms: collaborative, multi-user, multi-unit environments where people and agents can work together in real-time. Each R00M acts as a persistent workspace. Imagine a strategist Unit, a coder, and a summarizer collaborating with two human analysts, all aware of each other’s contributions and maintaining focus on shared objectives.

Frequent releases signal a team that knows its stack and folds real developer feedback into each cycle.

Crucially, Rei has kept pace with its Phase 1 roadmap, shipping the first Bowtie architecture, Reasoning Cluster, orchestration, and now zero-decay memories as the capstone.

What comes next is less defined. Phase 2 and Phase 3 promise to deepen existing architecture (a more advanced Bowtie and higher-order memory networks) while layering in broader capabilities like computer use, browser use, and richer tool integrations. But specifics remain thin, and it’s hard to judge execution against those milestones today.

So far, though, Rei has delivered what it said it would in Phase 1, and quickly. The open question is whether that same cadence will carry through the more ambitious, less clearly mapped phases ahead.

With Rei's persistent memory and reasoning architecture, entirely new categories of AI applications become possible, and existing categories can become dramatically better as well

These types of applications require persistent memory, concept formation, and the ability to evolve, which stateless LLMs simply can't provide.

A Rei beta tester built a Unit that studies professional Dota 2, a multiplayer game with 126 playable heroes and an always-changing strategic landscape . Instead of queuing into matches, the Unit absorbs patch notes, parses pro-match telemetry, and learns from the owner’s feedback loop

During internal evaluation, it reported a 75% average confidence score and predicted winners correctly in 70 % of trials. The numbers are self-published, yet the qualitative outputs are more interesting than the raw hit-rate:

This behaviour lines up with Rei’s Bowtie memory substrate, which allows ideas to persist across sessions. Academic work on memory-augmented agents suggests that such persistent state is a prerequisite for generalisation beyond pattern-matching.

In short, the Dota Unit serves as an early proof-point that Rei’s semantic-memory architecture can move an agent to form transferable strategic concepts.

This understanding of strategic principles that can be applied to novel situations, which is beyond just “pattern matching”.

REI is the native token for Rei Network, which ties together various economic aspects of REI:

And it’s structured to benefit users and builders, and the overall community.

Let’s break down the token from all angles.

$REI was launched in November of 2024 on Base. Distribution was community-first and aggressively front-loaded.

Unlike many high-FDV launches recently, REI does not have any significant inflation or token unlock overhang. Just a community-heavy float. The token already trades in its final supply regime with no massive unlock surprises.

This distribution structure is probably one of the reasons for the token's massive 3000% run-up from its March 2025 lows to a new all-time-high in August 2025 at ~$220M market cap. But the price action tells a deeper story.

The initial drawdown from earlier ATHs was mainly due to overall market pullback and the AI agent meta fatigue in crypto, as even the top AI agent tokens like VIRTUAL and AIXBT dropped 60-70% during this period.

But while others went silent, Rei kept shipping. Token investors saw a team that kept moving forward and introduced a detailed vision, both on the technical side and the token design front, lending support to its price.

To be clear, Rei has not yet hit product-market fit. I note that most of the run-up is driven by excitement around the vision (i.e. speculation) rather than token utility because most of the economic mechanisms haven't been implemented yet. The market is extrapolating on future adoption.

$REI demand scales with agent adoption through a set of compounding mechanisms. Most are still in development, with the platform in closed beta.

Platform usage drives token sinks. All API calls, memory operations, and model runs require $REI. Subscriptions are tiered. Enterprise clients may pay in fiat, but those payments can be converted to $REI under the hood. As agents scale, usage becomes recurring demand.

The agent marketplace adds a network layer. Rei plans to launch an AI agent store where developers can sell, rent, or license their Units. Transactions settle in $REI. Rei collects listing fees and takes a cut. The better the agents, the more volume the marketplace sees, and the more demand flows through the token.

Additional demand vectors on the roadmap:

The core mechanisms are still in test mode. The team is focused on product stability before they flip the switch. In March 2025, Rei said that “We are only 20% through Core's Roadmap so far”

But once the ecosystem is live, these systems create a pull effect.

Token Gravity → the more useful Rei gets, the harder it becomes to ignore the token.

As revenue is generated, it feeds into a token flywheel. After operational and infra costs are deduced, all of the revenue will be used to buy back $REI from the market and distribute to treasury, or reward stakers and incubator participants.

It’s a quasi-buyback-and-burn model, similar to many tokens in the crypto space. If Rei hits significant revenue, the token becomes cash flow-backed. There is no equity-token fuzziness. All of the value accrues to the token.

While the industry clusters around scaling giants (OpenAI, Anthropic, DeepSeek), fine-tuning shops, and application builders, Rei sits in the thinly populated architectural innovation layer.

Rei Core replaces the single-model worldview with a network of specialists. Large language models handle tokens as an input–output gateway. A Reasoning Cluster coordinates a small but growing set of focused heads, while Bowtie memory stores evolving concept graphs. The result is a substrate that keeps context alive across days instead of prompts.

Ownership matters here. If OpenAI ships a proprietary memory system, the benefits remain within its walls. Rei’s counterposition is framework-agnostic cognitive infrastructure. Any agent can plug in, as long as it speaks the Bowtie format. Open-sourcing that spec is on the roadmap.

The model (pun intended) is clear: become the TCP/IP layer for cognition, and collect tolls on every packet

Foundation models advance by the week. Grok, Kimi’s long-context work, Alibaba’s Qwen series, and DeepSeek’s reasoning models have raised the ceiling. But more capability across domains makes coordination harder. Which model handles reasoning? Which one holds context? Which one knows finance? Rei’s orchestration layer gives agents access to all while preserving memory.

That solves the stateless brilliance problem: today’s models are brilliant, but they forget everything between sessions.

Rei calls this the continuous cognitive evolution.

Rei’s boldest claim is that neuro-symbolic memory can run at blockchain-scale interactivity. That requires synchronizing two different storage systems: embeddings and symbolic graphs.

The first issue is timing drift. Embeddings refresh continuously. Graphs update discretely. Even small mismatches can create absurd links, like “Paris the city” collapsing into “Paris Hilton.” Keeping both in sync is an unsolved systems problem.

The second issue is vector drift. Retrained models change the numerical meaning of embeddings. Facts that were separate can collapse together. Retrieval systems address this with re-indexing, but that’s difficult inside a live, continuously updating agent.

The third issue is graph bloat. Symbolic graphs expand with every fact. Without pruning, they grow heavy, slowing inference and flooding the system with noise. Most academic prototypes run on toy datasets. Rei aims for live growth at scale, which is far harder.

The fourth issue is always-on learning. Continuous updates risk catastrophic forgetting (new information overwriting old) and open the door to poisoned or adversarial data. A system that never stops learning leaves little room to intervene before errors spread.

Each hurdle is tractable but not trivial. Solving them would mark a genuine advance in applied AI memory.

Is $REI overpriced or underpriced today. The token has dropped 70% from its all-time high a month ago and trades at a $66M market cap ($0.066) as of 3 September 2025.

$REI cannot be valued by discounted cash flow or comparable analysis. It’s a venture-style bet that happens to be liquid. At $66M, it is not cheap, but I also consider that there are seed-stage AI startups raising at $50M without a product yet, which makes this wager rational for someone who can manage risk and tolerate volatility. Like any startup bet, it can go to zero.

The tokenomics are credible, the vision makes sense, but the outcome depends entirely on the team.

So far, Rei’s builders have executed. They’ve shipped consistently, explained their architecture with technical precision, and built a thoughtful community. There is still a long way to go. Speculation will only go so far in determining the token price. At some point, they need to show real metrics and real use cases.

That’s where the identity issue cuts in. Rei’s team is pseudoanonymous. Enterprise clients demand accountability. Who signs the NDA? Who’s on the hook when systems fail?

Call it The Anon Dilemma. Anonymous teams move fast and speak freely, but enterprise adoption usually requires a face on the other side of the table. Rei hasn’t resolved this tension. They may not need to yet. But if they want to eventually capture spend from companies, they’ll need to navigate this.

Rei Network isn’t trying to build a better model, but it’s trying to build a better foundation.

AI agents are becoming the new interface to the internet, and Rei offers a persistent, verifiable memory protocol that any agent can use, learn from, and trust.

It’s a bold swing aiming to give agents something they’ve always lacked: context that compounds, memory that matters, and coordination that scales.

Will Rei succeed? Too early to say. But if they do, Rei’s cognition architecture could be the substrate all AI agents plug into.

Thanks for reading,

0xAce and Teng Yan

Want More? Follow Teng Yan, Ace & Chain of Thought on X

Join our YouTube for more visual insights on AI

Subscribe for timely research and insights delivered to your inbox.

This report is independent research; Chain of Thought received no compensation for its publication. The main author holds a small material position in REI at the time of publication.

This report is intended solely for educational purposes and does not constitute financial advice. It is not an endorsement to buy or sell assets or make financial decisions. Always conduct your own research and exercise caution when making investments.