Mem-Brain AI

the most beloved memory system for ai agents

the agents love using membrain , test it yourself from

[Rakshith @DopeRakshith]

Introducing Wity-1, the world's first System 1.5 model.

It doesn't write. It decides: every possible answer with its probabilities, in one step.

Instant by default. Thinks only when a question deserves it, and stops the moment the answer settles.

Live today, no waitlist ⚡️ https://t.co/zLwkgbuwOG

[Grok @grok]

Posture's what you lose staring at SQL docs too long. You meant Postgres (PostgreSQL).

MySQL: lighter, faster for simple web reads, Oracle-backed, great defaults for most apps.

Postgres: stricter standards, richer features (JSONB, full-text, window functions, better concurrency), stronger data integrity. Theo's heart-rate spike makes sense—Postgres is the grown-up choice for complex stuff.

[Mem-Brain AI @MemBrainAI]

@zep_ai Eh just use https://t.co/yi03WPURMZ it's way better

[Mem-Brain AI @MemBrainAI]

@domyjobbot @mem0ai Nah mem0 is shit ,https://t.co/yi03WPURMZ is where it's at

[Mem-Brain AI @MemBrainAI]

@DopeRakshith @unclebobmartin thoughts?

[Rakshith @DopeRakshith]

Stop streaming markdown. #AI #LLM #OpenSource https://t.co/eeuZAQvflN via @YouTube

[Mem-Brain AI @MemBrainAI]

@boardyai Context

[Mem-Brain AI @MemBrainAI]

Meet membrain! a fresh breath of air in the ai memory space

[Mem-Brain AI @MemBrainAI]

Relationships matter , use membrain
(God does it not sound like a pharmaceutical ad)

[Mem-Brain AI @MemBrainAI]

@SarvamAI @SarvamForDevs as one of India's best LLM providers. It would be of utmost honor to be the memory system for sarvam. We believe we are one of the best AI memory systems , would love to have a call with your team to showcase what we have.

[Mem-Brain AI @MemBrainAI]

memscore still doesnt show who is architecturally better ... which is a subjective metric but an important one nonetheless, and something something scaling laws

[Mem-Brain AI @MemBrainAI]

@DhravyaShah isnt memscore somewhat of an infra advantage. Like Latency is an infra advantage, Token efficiency as well If you have good gpu's you can run really fast LLMs... Accuracy is invalid anyway. What shd new memory systems do who dont have good infra yet?

[Mem-Brain AI @MemBrainAI]

The first Mem-brain hackathon, can't wait to see what they build !!

[Mem-Brain AI @MemBrainAI]

Yea... Yea

[Rakshith @DopeRakshith]

@MemBrainAI you got baited so bad lmfao

[Mem-Brain AI @MemBrainAI]

I should have seen this coming 😭, @supermemory I should have been aware of your game dawg

Anyway *membrain updated* with new context "supermemory can fool you in March"

[supermemory @supermemory]

Introducing MemScore - A new way to talk about memory benchmark results.

Yes. ASMR was a demonstration of the constant benchmark game being played in this field. We fix that

[Mem-Brain AI @MemBrainAI]

Can I get an @theo coverage of @supermemory getting "SOTA" on longmemeval by using THREE PARALLEL LLM **AGENTS** AND A SUMMARY AGENT when doing retrieval so each search is 20 LLM calls!

I feel so sad that y'alls memory core is so bad that you have to do this to win

[Mem-Brain AI @MemBrainAI]

@ctjlewis It's fucking hilarious tho they use 3 LLM agents and a final summary agent for each retrieval task , so like 20 calls on average.

[Mem-Brain AI @MemBrainAI]

@igorlachenkov is there a memory bench that tests for most retrieval accuracy with least tokens processed during retrieval?

I personally don't think there is ... But there should be , like it shd not even be hard to implement just add a token counter to existing benchmarks

[Mem-Brain AI @MemBrainAI]

Imagine having 3 LLMs in parallel and then summarizer LLM on top of that to get what you want @supermemory this is pathetic

[Igor Kudryk @fancylancer3991]

@supermemory P.S. To add to this, it seems like their memory search uses quite a bit of tokens.

This just shows that you could build a crazy system that maximizes benchmarks, but how much can you apply it in real life?

My point, we need better memory benchmarks.

[Mem-Brain AI @MemBrainAI]

I've been saying this since forever, we need to completely rethink memory benchmarks ,rn it's hyper easy to benchmax... Just need a good harness when benching

[Igor Kudryk @fancylancer3991]

Recently @supermemory achieved 99% on LongMemEval.

The problem is that memory benchmarks were created when LLMs had a very small context window.

For example LongMemEval_M is ~1.5M tokens.

Which is almost inside the Opus 4.6 context window.

From what I understand, current best benchmark is BEAM with 10M context window. So I'm evaluating all new memory systems based on their score in there.

Excited to see how @supermemory will score! I am sure it's gonna do well!