Taranjeet(@taranjeetio)

Chamath is right: being locked into one foundation model is quickly becoming a bottleneck for enterp...

Taranjeet (@taranjeetio) — Fri, 24 Apr 2026 21:21:51 GMT

Chamath is right: being locked into one foundation model is quickly becoming a bottleneck for enterprises.

But there’s another lock-in forming one layer deeper: memory lock-in.
It’s not just models that need to be swappable. The context that makes agents useful needs to be portable too.

When I switch from Claude Code to Codex, or from one agent framework to another, I shouldn’t lose the accumulated context that makes the agent useful in the first place: preferences, feedback loops, codebase decisions, prior tasks, and company-specific knowledge.

Otherwise, the model swap is technically easy but operationally expensive.

A serious enterprise AI control plane won’t just manage prompts, evals, requirements, tests, and feedback. It will also manage the memory layer underneath them.

Models should be interchangeable.
Context should be portable.

Memory should travel with the user, the team, and the agent across models, frameworks, and apps.

Model portability gives enterprises flexibility. Memory portability gives agents continuity.

This is why we started @mem0ai. We believe memory should be portable infrastructure for AI agents, independent of the model, framework, or app.

As enterprises adopt more agents and switch between more models, the question won’t just be: “Can I change the model?”

It will be: “Can my context come with me?”

The model can change. The context should persist.

That’s what we’re building at Mem0.

Chamath Palihapitiya@chamath

This is increasingly becoming a very useful feature for Enterprises as they a) decide to swap models because of cost b) decide to swap models because of one's ToS c) decide to swap models because of security d) decide to swap models because of capability etc etc etc

Being tied to one Foundational Model is quickly becoming a bottleneck that can and should be avoided if you value your independence and flexibility.

Software Factory resolves this for complex projects within Enterprises. A very useful, under-repped feature as it turns out...

💬3🔄0❤️21👀2765📊6

It is interesting that the headline here is not just what the model can do, but how many tokens it t...

Taranjeet (@taranjeetio) — Fri, 24 Apr 2026 04:16:38 GMT

It is interesting that the headline here is not just what the model can do, but how many tokens it takes to do it.

For a while, model capability mostly meant: can the model handle the task at all? Progress looked like bigger models, bigger context windows, more tokens, and more compute. That made sense when the main problem was simply getting useful output in the first place.

But once models become broadly useful, the evaluation changes. Cost and latency stop being secondary concerns and start becoming part of the capability signal.

A model that gets to the right answer in 500 tokens is meaningfully better than one that circles for 5,000 tokens before landing in the same place. The final answer may be similar, but the path matters. At scale, that path shows up in cost, latency, reliability, throughput, and user experience.

So the better question is no longer: did the model get the task right? It is becoming more and more: how cleanly did it get the task right under a fixed budget of tokens, time, and context?

That is one way to read “gets what to do” points to. The model reads the task correctly, uses the right tool, uses the right context, ignores what does not matter, and avoids unnecessary rambling.

Part of this is the model itself becoming sharper. But part of it is also the system around the model: better harnesses, better context management, and better taste about what to include or ignore.

This is why accuracy, cost, and latency should be evaluated and measured together as a combined measure of capability. The next important metric is going to be capability per token: how much useful work a model can do for a given budget of context, cost and time.

The new and upcoming wave of models, benchmarks and architectures will be judged this way. It's going to be better outcomes with fewer wasted tokens, for everyone.

At @mem0ai, this is how we already think about memory systems and benchmarks. Our new algorithm is a step toward making capability per token a standard way to evaluate memory in production.

Sam Altman@sama

It is smart and fast; per-token speed matches 5.4 and it uses significantly fewer tokens per task.

In my experience, it "gets what to do".

Rolling out today in ChatGPT and Codex. We are working with API customers on security and safeguards and plan to launch in API very soon.

💬1🔄3❤️22👀2115📊4

This is what token-efficient memory looks like @mem0ai You can now run agent memory that is cost-e...

Taranjeet (@taranjeetio) — Thu, 16 Apr 2026 18:08:20 GMT

This is what token-efficient memory looks like @mem0ai 

You can now run agent memory that is cost-efficient at production scale without sacrificing quality.

Open source

mem0@mem0ai

x.com/i/article/2044…

💬3🔄2❤️18👀1702📊6

Your @openclaw experience just got better with @mem0ai plugin

Taranjeet (@taranjeetio) — Thu, 09 Apr 2026 15:33:26 GMT

Your @openclaw experience just got better with @mem0ai plugin

mem0@mem0ai

How to add long-term memory to your @openclaw in a minute 🦀

tell your claw "setup Mem0 from mem0 .ai/claw-setup"

What’s new with Mem0 OpenClaw Plugin v1.0.4:
- 8 built-in tools for memory configuration
- Session-based + persistent long-term memory
- Multi-agent memory support
More 👇

Your browser does not support the video tag.

💬3🔄3❤️21👀3162📊6

Your agent can now add memory with @mem0ai CLI Agent-first and seamlessly integrated with your stack...

Taranjeet (@taranjeetio) — Wed, 08 Apr 2026 19:19:17 GMT

Your agent can now add memory with @mem0ai CLI
Agent-first and seamlessly integrated with your stack.

mem0@mem0ai

x.com/i/article/2041…

💬0🔄4❤️22👀2327📊4

https://t.co/gMkwF7RBaD

Taranjeet (@taranjeetio) — Mon, 06 Apr 2026 15:04:13 GMT

mem0.ai/cli

💬0🔄0❤️7👀880📊2

Introducing @mem0ai CLI 🚀 Give your agents memory from the terminal - add, search, and retrieve me...

Taranjeet (@taranjeetio) — Mon, 06 Apr 2026 15:01:52 GMT

Introducing @mem0ai CLI 🚀

Give your agents memory from the terminal
- add, search, and retrieve memory
- remember agents + developers's preference
- agent-first, JSON schema-driven

We built mem0 CLI to help agents remember

Get started
npm install -g @mem0/cli

Open source 👇

Your browser does not support the video tag.

💬13🔄23❤️191👀20642📊71

File system is observability for AI agents & humans.

Taranjeet (@taranjeetio) — Sat, 04 Apr 2026 04:33:21 GMT

File system is observability for AI agents & humans.

Andrej Karpathy@karpathy

LLM Knowledge Bases

Something I'm finding very useful recently: using LLMs to build personal knowledge bases for various topics of research interest. In this way, a large fraction of my recent token throughput is going less into manipulating code, and more into manipulating knowledge (stored as markdown and images). The latest LLMs are quite good at it. So:

Data ingest:
I index source documents (articles, papers, repos, datasets, images, etc.) into a raw/ directory, then I use an LLM to incrementally "compile" a wiki, which is just a collection of .md files in a directory structure. The wiki includes summaries of all the data in raw/, backlinks, and then it categorizes data into concepts, writes articles for them, and links them all. To convert web articles into .md files I like to use the Obsidian Web Clipper extension, and then I also use a hotkey to download all the related images to local so that my LLM can easily reference them.

IDE:
I use Obsidian as the IDE "frontend" where I can view the raw data, the the compiled wiki, and the derived visualizations. Important to note that the LLM writes and maintains all of the data of the wiki, I rarely touch it directly. I've played with a few Obsidian plugins to render and view data in other ways (e.g. Marp for slides).

Q&A:
Where things get interesting is that once your wiki is big enough (e.g. mine on some recent research is ~100 articles and ~400K words), you can ask your LLM agent all kinds of complex questions against the wiki, and it will go off, research the answers, etc. I thought I had to reach for fancy RAG, but the LLM has been pretty good about auto-maintaining index files and brief summaries of all the documents and it reads all the important related data fairly easily at this ~small scale.

Output:
Instead of getting answers in text/terminal, I like to have it render markdown files for me, or slide shows (Marp format), or matplotlib images, all of which I then view again in Obsidian. You can imagine many other visual output formats depending on the query. Often, I end up "filing" the outputs back into the wiki to enhance it for further queries. So my own explorations and queries always "add up" in the knowledge base.

Linting:
I've run some LLM "health checks" over the wiki to e.g. find inconsistent data, impute missing data (with web searchers), find interesting connections for new article candidates, etc., to incrementally clean up the wiki and enhance its overall data integrity. The LLMs are quite good at suggesting further questions to ask and look into.

Extra tools:
I find myself developing additional tools to process the data, e.g. I vibe coded a small and naive search engine over the wiki, which I both use directly (in a web ui), but more often I want to hand it off to an LLM via CLI as a tool for larger queries. 

Further explorations:
As the repo grows, the natural desire is to also think about synthetic data generation + finetuning to have your LLM "know" the data in its weights instead of just context windows.

TLDR: raw data from a given number of sources is collected, then compiled by an LLM into a .md wiki, then operated on by various CLIs by the LLM to do Q&A and to incrementally enhance the wiki, and all of it viewable in Obsidian. You rarely ever write or edit the wiki manually, it's the domain of the LLM. I think there is room here for an incredible new product instead of a hacky collection of scripts.

💬2🔄2❤️33👀5098📊6

Now you can add @mem0ai as the memory layer for your Hermes agent @NousResearch

Taranjeet (@taranjeetio) — Fri, 03 Apr 2026 22:08:10 GMT

Now you can add @mem0ai as the memory layer for your Hermes agent @NousResearch

mem0@mem0ai

x.com/i/article/2039…

💬1🔄1❤️36👀5359📊8

If you're using claude code @claudeai, you need to know this. - 200 lines memory limit in the sour...

Taranjeet (@taranjeetio) — Tue, 31 Mar 2026 21:03:40 GMT

If you're using claude code @claudeai, you need to know this.  
- 200 lines memory limit in the source code: MAX_ENTRYPOINT_LINES 
- 1 week into a project. 100+ memories saved. Index truncates. Old context gone. Claude forgets. 
- Memory is split into 4 types: user / feedback / project / reference.  

And the surprising part: every turn triggers a separate Sonnet API call just to retrieve memory.  

The fix: install @mem0ai Claude Code plugin.

mem0@mem0ai

x.com/i/article/2039…

💬2🔄2❤️40👀6322📊6

BEAM is a good starting point to benchmark the memory systems in the right way.

Taranjeet (@taranjeetio) — Sat, 28 Mar 2026 16:40:16 GMT

BEAM is a good starting point to benchmark the memory systems in the right way.

mem0@mem0ai

x.com/i/article/2037…

💬6🔄5❤️30👀5122📊12

Bytedance’s deerflow has been trending #1 on github. Read more to learn how the agentic memory works...

Taranjeet (@taranjeetio) — Fri, 27 Mar 2026 15:08:00 GMT

Bytedance’s deerflow has been trending #1 on github. Read more to learn how the agentic memory works.

mem0@mem0ai

x.com/i/article/2032…

💬4🔄2❤️37👀8972📊13

At @mem0ai - Claude Code Builders Lunch Thank you for coming!

Taranjeet (@taranjeetio) — Thu, 19 Mar 2026 20:15:14 GMT

At @mem0ai - Claude Code Builders Lunch

Thank you for coming!

mem0@mem0ai

Great vibes at Claude Code SF Builders Meetup @ mem0 HQ

💬3🔄0❤️18👀1024📊4

. @mem0ai also.

Taranjeet (@taranjeetio) — Mon, 16 Mar 2026 04:28:22 GMT

. @mem0ai also.

Karri Saarinen@karrisaarinen

At this point, it feels like most teams building great products are on @linear (even if not mentioned here, they're probably on or currently trialing):
Polymarket
Perplexity
Supabase
Cash App
Coinbase
Substack
Mercury
Raycast
Lovable
OpenAI
Cursor
Vercel
Replit
Ramp
Boom
Brex
Clay

💬0🔄0❤️3👀389

Welcome @ellen_in_sf to the team. Super excited to be collaborating with you. More coming 🚀

Taranjeet (@taranjeetio) — Thu, 12 Mar 2026 00:36:40 GMT

Welcome @ellen_in_sf to the team.
Super excited to be collaborating with you.

More coming 🚀

ellen ᯅ 🇺🇸🇮🇩@ellen_in_sf

just got a new launched @Apple macbook m5 max on my new role at @mem0ai 

Thanks @taranjeetio 

should i gave full control to run @openclaw or run 20 claude code terminal to claudemax my work

💬0🔄0❤️29👀2515📊3

Every agent need email. Congrats @agentmail team.

Taranjeet (@taranjeetio) — Tue, 10 Mar 2026 17:26:33 GMT

Every agent need email. Congrats @agentmail team.

💬3🔄0❤️11👀707📊5

This one is going to be great one. Do attend if you can.

Taranjeet (@taranjeetio) — Mon, 02 Mar 2026 17:45:26 GMT

This one is going to be great one.
Do attend if you can.

Daytona@daytonaio

The infrastructure layer is being rebuilt for AI, agents, and a new kind of cloud.

So we’re hosting our inaugural @Daytonaio COMPUTE Conference on the Warriors’ home court, bringing 1,500 builders together.

March 9, 2026

Apply for tickets below

💬1🔄0❤️7👀1240📊2

We just crossed 100k developers using Mem0 What began as an experiment to solve AI amnesia problem ...

Taranjeet (@taranjeetio) — Thu, 22 Jan 2026 18:32:33 GMT

We just crossed 100k developers using Mem0

What began as an experiment to solve AI amnesia problem is now used by thousands of developers building AI agents.

Grateful to everyone building and pushing us to make it better every day.

Still early. A lot more to build.

💬8🔄3❤️47👀2695📊15

I liked how Sam Altman framed memory. What stood out was how early this still is. Most AI agents t...

Taranjeet (@taranjeetio) — Thu, 08 Jan 2026 16:56:41 GMT

I liked how Sam Altman framed memory.

What stood out was how early this still is.

Most AI agents today are still largely stateless. They perform well within a session, but once it ends, context resets.

You can improve models and add more tools, but without persistent state, agents don’t improve with use.

This is not a small gap but a structural one.

That’s why memory is becoming a core part of agent architecture.

Not just storing facts, but persisting:
- prior decisions
- user preferences
- distilled context from past interactions

This is also where personalization actually comes from, systems that adapt over time and behave consistently because of accumulated state.

This perspective is what’s shaped how we think about memory at @mem0ai.

Your browser does not support the video tag.

💬7🔄7❤️73👀6626📊25

Hiring our first data hire at @mem0ai. This isn’t a role where you join a data team, you’ll build i...

Taranjeet (@taranjeetio) — Wed, 07 Jan 2026 05:39:58 GMT

Hiring our first data hire at @mem0ai.

This isn’t a role where you join a data team, you’ll build it.

From day one, you’ll own pipelines, metrics, dashboards, and what actually gets measured.

Mem0 is the memory layer for AI agents. YC-backed, $24M raised, 45k+ GitHub stars, used by developers at some of the largest companies globally. We’re past “will this work?” and deep into scaling.

Looking for #hiring someone who’s done this at a startup before. Strong opinions, high ownership, excited to build from scratch.

📍Bangalore (hybrid)

Apply from the link below👇

💬17🔄8❤️260👀21738📊66

https://t.co/f8oas7UEXi

Taranjeet (@taranjeetio) — Wed, 07 Jan 2026 05:39:58 GMT

mem0.dev/careers/data-a…

💬3🔄0❤️10👀1948📊5

2025 memories wrapped: - Released our first Mem0 research paper - Launched OpenMemory MCP - Shipped...

Taranjeet (@taranjeetio) — Sat, 03 Jan 2026 07:59:33 GMT

2025 memories wrapped:

- Released our first Mem0 research paper
- Launched OpenMemory MCP
- Shipped the OpenMemory Chrome Extension
- Crossed 44k+ GitHub stars
- Hit 18M+ Python package downloads
- Powered production workloads at fast-growing startups and Fortune 500 companies
- Integrated Mem0 natively into major agentic products like CrewAI, Flowise, and Langflow
- AWS selected Mem0 as the exclusive memory provider for their new Agent SDK
- Closed a $24M funding round
- Our co-founder & CTO @deshrajdry presented our research at ECAI 2025 in Bologna
- Met the team in person for our @mem0ai offsite in Goa, India

That’s 2025. Grateful for everyone building with us. 2026 is going to be even bigger.

💬3🔄3❤️79👀8878📊14

Applied to YC 7 times before finally getting in. Here’s what I learned the hard way👇 When I first ...

Taranjeet (@taranjeetio) — Mon, 03 Nov 2025 17:27:41 GMT

Applied to YC 7 times before finally getting in.
Here’s what I learned the hard way👇

When I first started applying to @ycombinator, I treated the application like a pitch deck, trying to convince YC to invest, instead of explaining why the problem mattered.

Over time (and 7 applications + 3 interviews later), I realized something simple but powerful "the best YC applications read like honest conversations, not investor pitches".

The application itself is a forcing function, it makes you clarify:
- What’s the problem and who’s the user?
- How big is the opportunity?
- Why are you the best team to solve it?

If you’re struggling to explain these clearly, it usually means you need to go deeper into your own understanding of the problem.

In my last interview, I opened with a short framework:
one line on the product, the problem we’re solving, why it matters, who it matters for, our traction, and how big it can get.

That 60-second summary changed everything. The rest of the conversation flowed naturally.

My advice to anyone applying to YC: Prioritize clarity over complexity. The simpler you explain your idea, the stronger your conviction comes across.

If you’re applying this batch and want feedback on your application or if you’re building something that needs memory, my DMs are open.

💬30🔄35❤️628👀61693📊245

Memory is what makes us human. It's also what makes AI truly intelligent. @mem0ai has raised $24M t...

Taranjeet (@taranjeetio) — Tue, 28 Oct 2025 15:33:09 GMT

Memory is what makes us human. It's also what makes AI truly intelligent.

@mem0ai has raised $24M to build the universal memory layer for AI.

Thousands of teams in production. 14M downloads. 41K GitHub stars.

Intelligence needs memory & we're building it for everyone.

More👇

Your browser does not support the video tag.