This is the actual bottleneck. The models are smart enough already. What is missing is the company-specific context locked in senior people heads. Whoever cracks knowledge extraction at the company level unlocks the rest.
As you work on this, please consider using GBrain as your OSS retrieval layer
https://t.co/0F5uDQzPHu
#RFP Submission:
Clude Labs partnering with {REDACTED} to fine-tune the first memory-specific model
This grant proposal outlines a joint research initiative between Clude Labs and {REDACTED} to prototype and evaluate the first production-grade memory-tuned language model architecture.
Research on memory-conditioned models has existed for years. The core insight is straightforward: model weights encode persistent behavior, preferences, and personality more effectively than prompts or retrieval systems ever will. Despite this, no commercially deployed system today performs per-user memory fine-tuning at scale.
The primary blocker has not been compute infrastructure. It has been training data quality.
Fine-tuning directly on raw conversational history consistently produces degraded outputs. Contradictory user statements reinforce conflicting behaviors, stale preferences remain embedded indefinitely, and high-noise interaction logs dilute meaningful long-term signal. Existing approaches to personalized model adaptation have therefore remained largely confined to research environments, where curated datasets could be controlled manually.
Clude Labs approaches the problem differently.
Our memory architecture was designed from inception for distillation and consolidation rather than naive retrieval. The system already classifies memories by type, including durable facts, behavioral patterns, preferences, decisions, and contextual states. It continuously evaluates recency, confidence, conflict resolution, and relative importance across memories over time.
This curation layer transforms memory from an unstructured chat archive into a high-signal training substrate suitable for model adaptation.
On top of this infrastructure, we propose training per-user LoRA adapters: lightweight low-rank parameter deltas applied to a frozen base model. Rather than retraining the foundation model itself, individualized adapters encode user-specific behavioral priors and long-term personalization while maintaining efficient inference and serving costs.
The proposed research program focuses on four core objectives:
1️⃣ Training Pipeline Validation
Develop and evaluate end-to-end pipelines for generating clean memory distillations suitable for continual adapter fine-tuning.
2️⃣ Dynamic LoRA Rank Selection
Investigate adaptive rank allocation strategies where sparse memory profiles utilize smaller adapters while users with dense, consolidated histories receive higher-capacity representations.
3️⃣ Training Cadence Optimization
Measure the effectiveness of nightly versus weekly adapter refresh cycles, including retention quality, behavioral consistency, and inference stability.
4️⃣ Production-Scale Serving Architecture
Prototype low-latency adapter loading and routing infrastructure capable of supporting real-user deployment volumes.
The central hypothesis is that curated memory fine-tuning will outperform retrieval-only personalization systems across retention, coherence, and long-term behavioral alignment metrics.
If validated, this architecture represents a meaningful shift in how personalization is implemented in AI systems. Instead of models repeatedly retrieving fragmented context about a user, the model itself incrementally adapts over time through persistent individualized weight updates.
The end state is not simply an assistant that remembers information about a user.
It is a model that gradually becomes shaped by them.
In light of recent events across the crypto AI startup landscape, we have decided to reinforce our long term commitment by extending the lock on 7% of team supply for another 2 months.
This brings total locked supply to almost 20%.
The trust from our community and holders is not something we take lightly.
Words matter.
Actions matter more.
Our lock schedules are intentional and aligned with what we are building toward. They are structured around long term growth objectives. We will share more once negotiations are complete.
Clude is now cashflow positive through enterprise revenue. We are no longer dependent on trading fees for operational costs. This gives us flexibility to strategically allocate this for buybacks, which we have consistently done since launch.
Our vision remains unchanged:
1. Build PMP into the global standard for Tokenized Memory, starting with Singapore.
2. Turn memory into a true digital asset powering the future of Programmable Intelligence.
3. Dream big enough to create the stablecoin-like asset layer for intelligent agents.
We are here to build until this becomes reality.
Not for weeks.
Not for hype.
Memory is the moat.
Memory will become an asset.
🔐 Just locked 71,346,564 $Clude tokens with @Streamflow_Fi
It's on-chain. You can check the amount, time-period and recipients.
Check it out👇
https://t.co/UEoM7m4VeC
Busy day irl today with back-to-back to meetings and honestly I am still processing that I get to help define memory frameworks at this scale. We're working with regulators in Singapore right now to build a tokenized memory standard for agents. On-chain, immutable decision logs with full explainability.
Singapore wants to lead on responsible AI regulation, and being here to shape that standard with them is a pretty rare position to be in. Persistent, verifiable, tokenized memory is what closes the gap between frontier AI and the trillion-dollar industries that won't touch it without receipts.
Singapore isn't just talking about these, MAS has been running Project Guardian since 2022, a public-private collab on asset tokenization that's moved past pilots into real frameworks. They published guidance on tokenizing capital markets products, they're trialing tokenized MAS bills settled with wholesale CBDC this year, and in April they proposed lower capital requirements for banks holding tokenized assets on permissionless chains. Basically telling institutions "you can actually use this."
On the AI side, IMDA published a governance framework specifically for agentic AI in January, backed by AI Verify as the open-source testing toolkit
Singapore is one of the only places where the tokenization infrastructure and AI governance rails are being built at the same time, by the same regulators. That's the overlap we're building in.
Memory is the unlock
Been getting some feedback on exactly how and what @cludeproject does.
Hope this feature will shed some light but there's a ton more you can do so perhaps I could run this as a series!
Any constructive feedback or ideas - Please feel free to tag me
The @Solana Frontier Hackathon has concluded. Thanks to all builders who entered products!
We're organizing the submissions & preparing to initiate the judging process, which will take 5/6 weeks.
We will announce the total # of submissions & share a public directory soon.
🔐 Just locked 49,693,532 $Clude tokens with @Streamflow_Fi
It's on-chain. You can check the amount, time-period and recipients.
Check it out👇
https://t.co/lBGnrdbRt9
@ErikVoorhees Been solving this exact issue for the past couple of months. We're free to use and competitive with the best players out there. Also - We integrate with Venice too!
Two weeks ago, we set out to make AI memory portable.
Today, Clude is the first AI memory layer with a real, signed, open file format. Every memory carries a cryptographic receipt, content-hashed, signed with your private key, and anchored to Solana. Anyone can verify it. No one needs to trust us.
We shipped a tiny standalone verifier. Install a thirty-kilobyte package, point it at any pack, and audit it without a Clude account. Backups, encryption, attachments, signed soft-deletes for GDPR, all in the spec.
On top of that, we built the Brain Wiki. Your conversations don't disappear into chat history. They auto-organize into a structured, cross-linked knowledge base. Open questions. Decisions. To dos. The agent flags contradictions when older notes disagree with newer ones.
Install a vertical like Compliance, and memories about audits, SOC2, or regulator asks start routing into the right topics automatically, by keyword, and by embedding similarity. Export the whole thing as markdown, open it in Obsidian, or share a single topic with a colleague.
Now the bigger play: memory tokenization.
Every memory becomes a first-class digital asset. Your AI's knowledge, every conversation, every decision, every preference it learned about you, becomes something you can own, prove, and move.
You grant or revoke access cryptographically. You move memory between agents, Clude, ChatGPT, Claude, anything that supports the open spec. You prove what
your AI knew, and when, without trusting the vendor.
Memory marketplaces. Auditable AI. Multi-agent interop with cryptographically verifiable trust.
This is the moment memory stops being a feature buried inside someone else's product. It becomes a protocol. Open. Portable. Yours.
Welcome to the memory layer
Birdeye API access is coming to Clude users. Free.
Real-time onchain data piped into a memory layer that compounds over time. Your agents don't just see the market, they remember it.
Most AI trading agents forget everything after each session.
@cludeproject, one of 12 @Pumpfun BIP Hackathon winners, fixes that.
A persistent memory layer that tracks trades, context, and outcomes, then feeds it back into the agent. It gets smarter the more you use it.
Through our $50K infrastructure grant, Birdeye Data is now powering Clude to bring real-time onchain data directly to their users.
This is the exact mental model we've been executing at @cludeproject for the last couple of months.
We didn't just theorize about "experiential memory" and context harnesses: we shipped the on-chain memory layer that makes it real:
• Persistent, forkable, never-dies memory (literally immutable on Solana)
• Dream cycles + JEPA predictor that distill raw traces into higher-level primitives and emergent connections
• Sub-200ms recall with full provenance so every injected fragment is auditable and tamper-proof
• Native governance baked in from day one (Clude Compliance launching now)
The Bitter Lesson is playing out exactly as described. Manual engineering of prompts/hooks won't scale. Scalable search + immutable compute infrastructure will.
That's why we built Clude as the memory & accountability layer for agentic AI. Agents accumulate insane experiential data? We turn it into a verifiable, self-managing long-term system enterprises can actually trust.
KYA isn't optional anymore. It's the foundation.
🧠⛓️
When we launched Clude, people asked why we built on blockchain. "Just use a database."
We built on Solana because one thing had to be true from day one: immutability. Once an agent stores a memory, nobody can go back and edit it. Not us. Not the user. Not a compromised server. Nobody.
Solana was the obvious pick, fast enough to not bottleneck agent cognition, cheap enough to anchor every memory without burning through a treasury.
We kept saying KYA.. "Know Your Agent" would matter. That as agents go from toys to tools to autonomous workers, someone's gonna need to answer for what they do.
Since then we've been heads down on the memory layer. Building, testing, iterating. 5 memory types. Sub-200ms recall. Dream cycles that let agents consolidate, resolve contradictions, and actually evolve their understanding over time. It works. It's been proving itself in production.
But here's the thing most people missed, auditability wasn't a feature we added later. We baked it in from day one. Every memory stored, every recall, every dream phase: timestamped, traceable, verifiable on-chain. We didn't build memory and then figure out compliance. We built memory knowing compliance would be the unlock.
Now we're ready to scale. And the timing couldn't be better.
Right now there are thousands of agents being deployed into production every week. Accessing customer data. Making financial recommendations. Taking actions. And the entire audit trail for most of them is a prompt log in some S3 bucket that anyone with access can modify.
That's not an audit trail. That's a pinky promise.
Immutability isn't a feature. It's the foundation. And most people are building without one.
We've got plans. Sharing more soon over the coming weeks 🧠⛓️