cha

Building second brains for the AI native

24&7 w/ Hermes | Git | Model Agnostic

Bleeding Edge AI workflows integrated to your second brain

Waitlist incoming

[cha @chakhan22]

We are now the bottleneck.

AI agents can now work for hours. They can handle tasks of ever increasing complexity and do them correctly.

They do a task and come back. Wait for us to give them another one. But we already have another one running, and another one, and another one...

The key to unlocking the multiples of productivity gains with AI is in setting up systems that allow AI agents to do actual useful work without waiting for us after each prompt.

But how do you do that when so much of what AI produces is generic AI slop?

The answer is context. AI needs the correct context to do the job in a specific way, in your way.

To do outreach it needs your sales strategy, voice, tools, your history, an understanding of what works, what doesn't...

To do prospecting it needs your ICP, product, tools to find them, a feedback loop so it can evolve its strategy...

Your agents can only be so useful as much as they have the correct context when they need it.

This is not an easy problem.

Your context is ever changing. You get live data everyday, that needs to flow in, older stale data needs to flow out, learnings need to compound, decisions need to persist, older data needs to be pruned.

You crack this and you can have agents running an entire businesses' worth of operations for you while you sleep.

You can check in everyday, set directions, review outputs, develop new agents, and let the others cook.

Compound this over a year and the returns will be unimaginable.

I'm doing exactly this. Setting aside the hype, the 100x productivity gains, for the real, nitty gritty of getting my AI agents the real world context they need to do real, productive work for me while I sleep.

Follow me to see:
- the state of the art in memory systems
- how to build, grow, and organize your second brain on auto, without you needing to guide everything, and without it turning into AI slop
- methodologies of organizing your second brain so you can have AI agents find relevant context, and work

and how to iterate to make your systems more and more useful.

Cheers!

[cha @chakhan22]

Are you AI native?

Here's my self assessment.

Turns out, I am not. As defined by Boris Cherny, the creator of Claude Code, and a token-maxxing king.

Here's how he defines it:
Level 0: Gated - Agents cannot enter real work.
Level 1: Assisted - One live agent works beside the person.
Level 2: Parallel - Several independent workstreams run at once.
Level 3: Supervised autonomy - Repeated work runs without a live conversation.
Level 4: AI-native - The system finds, ranks, and delegates work from intent.

If you talk to multiple agents at once: You are level 2.

If your agents go and do things without you there, summon their own agents, and do entire workflows, you are level 3.

I'm somewhere in between.

Level 3 things my agents do:
- automatical PR + refactoring on code changes,
- tests run automatically on every code change,
- UI verification on auto by Codex when UI is updated
- some agents on loops: daily review, daily X checks for new updates, X reply agent to find posts to reply to, ...

I still initiate the sequence with most chats. Still level 2.

But this roadmap has me thinking. What do I need to do to move more towards Level 3?

I'm looking at it from that lens for my second brain.

What info can it capture such that it moves me towards level 3 by constantly improving my workflow?

What if your AI automatically captured:
- opportunities to create new automated workflows
- improvements to AGENTS.md/CLAUDE.md that remove consistent issues

Do you have anything like this in mind? Let me know if you do.

[cha @chakhan22]

@tom_doerr Nice. Building something similar.

[cha @chakhan22]

@TrianaKurtetti Creating a second brain. Easy.
Managing the knowledge over time so it's useful, not so much.

I'm building a product that integrates all best practices on this as they emerge.

Waitlist soon.

[cha @chakhan22]

@0xkkai Dario just told you yes :D

And he's using Obsidian. Hilarious.

[cha @chakhan22]

@RoundtableSpace Making a second brain out of notes is extremely simple. Making it useful, not so much.

Second brain is only as useful as it being used to accomplish goals, handle tasks, do work for you.

[cha @chakhan22]

Will do. Your approach sounds great. I'm focusing on solo founders and solo approach at least at the beginning to not deal with multi tenancy. But just cracking access control, who gets to read, and change what file and/or information, how do you store it etc. that problem is worth billions.

[cha @chakhan22]

@rayimbuilds Just starting off. I have it for me personally. Productizing it, collecting best practices all across to integrate into a single system to be used by anyone who wants to become AI native. Cooking the MVP atm. Will take a week or two probably.

[cha @chakhan22]

Your second brain needs raw data.

Each model update brings a leap in intelligence.

Each time, you can organize your knowledge base again to solve issues you previously had.

This is only the case if you have the raw data AND if you have a record of the problems.

If you capture things and keep them unchanged, future AI can solve these issues.

So simple lesson. Store the raw data. Never delete it.

Also, use git, commit always. Your second brain needs it.

[cha @chakhan22]

@DanielLozovsky Would love to find more about memory poisoning. You know anyone who's doing some good work on this?

[cha @chakhan22]

@0xVeyyy Sweet. Did you try yourself?

[cha @chakhan22]

Does your AI need memory?

- re-explaining the same thing again and again?
- can't find the thing from a session from 3 weeks ago?
- want your AI to just know things?

If you answered yes to any of these, then probably yes.

But what is AI memory?

At its simplest, AI memory is a system that

1. stores your context
2. retrieves it during prompt time,
3. injects them into your prompt

So each message to your AI already has information from before. Sounds great.

Here are some I've tried and my experiences:

claude-mem: automatically captures your prompts, saves key info and injects it back into your prompts. I used it for about 4 weeks but didn't notice it was gone after I deleted it.

agent-memory: Had issues with this one. It created 1 real memory, 2,162 junk "lessons". It spawned ~300 background Claude sessions and ate ~20M tokens in one afternoon. I uninstalled.

qmd: local hybrid search over your markdown (keyword + vector + on-device re-ranking). Hand it to your agents as a search tool and it might solve your find my X problems outright.

And some extras I am looking to play with in the future:

Open Brain: one database + vector search any AI plugs into.

mnemory: self-hosted MCP server, plug and play, your data on your infrastructure.

basicmemory: local-first Markdown knowledge base over MCP. Your agent reads and writes plain files you can open yourself.

The space is insanely deep. Each tool has a different approach and are good for different use cases.

Follow for more on how to make your AI know you. I post about agentic OS, second brains, and AI memory.

[cha @chakhan22]

@bcherny Which company actually has enough software development work that can feed off a fleet of 1000 agents to work per engineer 24/7?

[cha @chakhan22]

@JoshEdmondson_ @fivosaresti Great question.

[cha @chakhan22]

@gold_tesla @jimmyslagl Have you tried Jake Van Clief's methodology?

Clear separation of responsibility and standardized routing within your folders. I think the latest models can handle it no problem. Opus level too. As long as organization is clear and enforced.

https://t.co/raFt07N4dj

[cha @chakhan22]

I don't know about that. The model by itself doesn't know what is there to retrieve, unless you tell it through CLAUDE.md and other methods. It might get good at understanding what is important to pull in for that request, but unless it already has a good harness around it, it will not do that.

[cha @chakhan22]

@hello_ambi What is it though? Automatic context ingestion based on AI memory? Knowledge base?

[cha @chakhan22]

Follow for more on this space. I'll share details and complete workflows of how to actually build yourself an AI context system that knows you. And one that actually solves the problems beyond a simple 20 minute demo.

[cha @chakhan22]

You want your AI to know you. Here's how.

There are many names for this: a second brain, an agentic OS, an AI chief of staff, AI memory, simple files and folders...

Each is an approach, a tech stack, a different solution to the same problem.

None are yet complete. None are yet a Jarvis to your Iron Man. But they're pretty darn close.