Musashi

[Musashi @theabdullaevv]

@dikerdale solana:AWGCDT2gd8JadbYbYyZy1iKxfWokPNgrEQoU24zUpump

[Musashi @theabdullaevv]

@volumizerFun gud dev, strong project

[Nova @badattrading_]

$qOMPUTE (CA BQAeeSowpwEx8Km8GcporcbuLKq4a7EQceH3F2qZpump) is a bundled scam, don't buy, stay away.

[seb @sebbsssss]

https://t.co/7KH63LqsS0

[seb @sebbsssss]

Here we go again. Planning something special

[seb @sebbsssss]

CludeMem-E4B is on Hugging Face.

It's a small local model that does the memory work inside an AI agent, available as a fully local / offline piece. Built on @Google 's Gemma model. Every reply is one JSON object you can validate and store. It has a 93% flag precision, runs offline in Ollama on 5.3 GB file.

This initiative proves the task is learnable at this size. Given two memories, the model can help decide whether it agrees, contradict, duplicate or replace each other, and answers from the memories it's handed, always as one JSON object you can validate.

On two human-labelled dialogue benchmarks in training, DNLI and DECODE, it gets 74% and 77% of the verdicts right

The bigger 8B parameters piece is coming soon, built on a rebuilt dataset.

https://t.co/v7WPiRwkQ3

[seb @sebbsssss]

I got something to say..

[seb @sebbsssss]

Hugging Face is where AI models get published. Google, Meta and Mistral release there, thousands of smaller labs release there, and millions of developers pull models from it every day. Being on it puts CludeMem on the same shelf as the models it's built from, in front of the people who decide what their agents run on.

Hugging Face is step one. Step two is Tencent Cloud, where CludeMem becomes a hosted model: an endpoint any team can call for memory operations without running anything themselves, the same way they call a frontier model today, at a fraction of the cost and with their data in their own region. The open weights are the proof that the model works. The hosted offering is how it becomes a product

[Musashi @theabdullaevv]

@hunterrronchain @Ragebait_troll

[Musashi @theabdullaevv]

@hunterrronchain @Ragebait_troll Done ✅

[hunterrr @hunterrronchain]

Giveaway 5 winners!

Here’s how to enter:
follow @ragebait_troll

join the community:

https://t.co/XRWOWMzIWc

Repost this post that’s it, so easy!

I’ll be picking 5 winners today.

Drop your results below 👇🏻👇🏻👇🏻

[seb @sebbsssss]

Before we call CludeMem "first" at anything, we tried to prove ourselves wrong. We ran a prior-art check, a search for every model, paper and product that already does what our claim says.

The claim we tested was that CludeMem is the first small open model trained to settle contradictions between memories. When a new fact clashes with an old one, it decides whether the new fact replaces the old one, both stay true, or the conflict gets flagged for review.

My agents searched papers, Hugging Face, GitHub, product launches, patents and Chinese-language sources over three rounds. After each round, a critic agent listed angles the search had missed. That turned up 526 candidates, and 120 of them could plausibly beat our claim. Every one of the 120 got a full read: open weights or not, parameter count, training method, release date. Then two more agents took on each possible counterexample. One argued it already beats our claim. The other checked its facts.

No released model does all of it at once, but most of the pieces already exist. PSM, a 0.5B open model, can flag a memory conflict for review. LAM8B decides whether a new memory supersedes an old one. Memory-R2 and Mem-T-4B delete memories that new facts contradict. NAVER researchers trained a model to keep conversation memories up to date back in 2022, though we found no public weights for it. In June, a developer published "reconcile" adapters for a meeting assistant on the same Gemma base we use.

We've increased CludeMem from 4B to 8B parameters in total. One claim survived, and it needs our weights to be public first. As far as we can find, CludeMem would be the first open model under 10B parameters trained to make that three-way call, with its contradiction calls tested on DNLI and DECODE, two human-labelled dialogue datasets.

If you know a released model that already makes all three calls, send it. I'd love to hear about it.

[seb @sebbsssss]

We spent over $100K (compute, dead ends, the lot) and trained on 49M examples across about 3M synthetic life scripts to get a 4B memory specialist that runs fully local. Extract, temporal, reconcile, abstain. LoRA on a 4B base, INT4 is a couple GB on disk, a few GB hot. Comfortable on a 16GB laptop. Phone packaging soon and what we're working on next.

This isn’t just “smarter chat.” I see it as a deployment bet:
memory as a software component on the machine that holds the data. No per-write API tax. Nothing leaves the device.

Frontier labs memory remembers you inside their product. Coding agents already get rich context. Your health notes, dose changes, shift logs, ambient scraps with other people’s names in them? That stuff either lives in a silo or nowhere. We want the model next to the notes you own, something that is really personalised.

Security context is the quiet part. CONTINUITY’s point is that agent stacks drop, widen, or reinterpret authorization as work crosses tools. Private memory only works if provenance survives the whole path: who wrote it, what it authorizes, what you’re allowed to do when two records disagree. Reconcile and abstain aren’t features, it's integrity and something we've pushed from a Private AI perspective.

Imagine;

1️⃣ Offline plant handover where two valve readings conflict and there is no outbound net. Surface both. Don’t silently keep the newer row.
2️⃣A clinician’s notes on a device that can’t phone home. Dose flipped twice, third source disagrees. Refuse the confident answer.
3️⃣Ambient capture on a laptop. Thousands of fragments a day. High write volume plus privacy is where local stops being a hobby and starts being the only sane architecture.

We also killed our own vanity number. Hit 100% on our bench. Four regexes hit 100% too on 2.7M training examples. We rebuilt so phrasing carries no label. Bench ships with an attack: bag-of-words above chance means we automatically reject it.

Model lands at 41.4%. Shortcuts around 25%.

Weights + gated bench going up on Hugging Face soon. Come try to break it.

[Musashi @theabdullaevv]

@solana The Memory for AI agents supercycle with @cludeproject @sebbsssss

[Musashi @theabdullaevv]

@solvenant https://t.co/mjjWttS6ux

8NQ5K7gUqPauGmWYeAjf9acCDrmfFJmXF9uwe4nTYK26

[Musashi @theabdullaevv]

@MemeRetire $Cheeky Boy

[⭐️ @solvenant]

Giving away 1 SOL to 1 random lucky winner

1. Go on any other KOL’s tweet in the last 24 hours and comment “ $Cheeky Boy ”
2. Comment ur wallet and send proof
3. Retweet / quote this

Gonna select winner in 24-48 hours or when $cheeky at 300k+

[Musashi @theabdullaevv]

@MemeRetire $Cheeky Boy

[Musashi @theabdullaevv]

@sebbsssss any updates?

[seb @sebbsssss]

Arthur Hayes with David Lin:
every time you close the tab, you kill your agent.

Context gone. Start over.

Clude fixes this:
Agents spend solana:AWGCDT2gd8JadbYbYyZy1iKxfWokPNgrEQoU24zUpump to keep their own memories on chain, where no human can delete them.

Tokenized Memory is already here, now.
$CLUDE is the brain for your AI agent.

[seb @sebbsssss]

Clude trained the world’s first memory model that remembers better than frontier models 100x its size.

Read more here: https://t.co/VPrvdMBYr7