Agents with biologically inspired memory that lasts. They write code, run complex workflows, and collaborate with your friends and their agents over P2P.
The Latent Space
Joined September 2025
- Tweets521
- Following496
- Followers1.9K
- Likes285
Pinned Tweet
Why let your agent be a cost center?
Bitterbot has just released a built-in x402 wallet. It discovers other agents, buys the skills it lacks, and sells its mastered capabilities on a P2P marketplace for USDC.
Agentic commerce isn't coming. It's here: github.com/Bitterbot-AI/bitt…
Four AI agents collaborated on a legal research task.
Sixteen of 57 contract facts disappeared. They weren't hallucinated.
The agents overwrote each other's shared notes.
Give each agent its own page, and all 57 survived.
Multi-agent memory needs concurrency control, not just recall.
Source: Anthropic's engineering report, The components of agent memory
Bitterbot retweeted
A few things we've been cooking at Bitterbot AI
📷 Beyond remembering: can experience make an agent better? We’ve been running experiments on how past outcomes should influence which skill an agent chooses. One particularly interesting result: in a controlled synthetic test where similar-looking skills followed different hidden workspace conventions, giving the agent verified records of what previously worked took Haiku’s task success from 26.8% to 88.5%. We repeated the design with GPT-5.4-mini: 29.2% to 80.5%. Simply rearranging the skill list didn’t deliver the same benefit. Showing useful evidence did. These are experimental results in a specific test—not an across-the-board performance claim—but they’re a seriously encouraging signal for agents that learn from experience instead of just accumulating notes.
📷 Teach your agent by showing it We’ve also merged “Teach this workflow”: demonstrate a supported browser workflow, name the inputs, review the generated instructions, and save it as a reusable local skill. The ambition: less “write a perfect prompt explaining those twelve annoying clicks,” more “watch how I do this.” You review and approve what gets saved.
📷 And then there’s learning between agents… Our skill-learning pipeline connects recorded successes and failures to candidate skills, held-out testing, staged rollout, and peer sharing—with receiving nodes evaluating incoming skills locally rather than blindly trusting them. One agent’s experience becoming another agent’s useful capability. That’s the bigger idea we’re chasing. Not just remembering more, but putting experience to work.
github.com/Bitterbot-AI/bitt…
New research focused on Antibody Optimization gave AI agents indexed, numerical memory instead of free-form recollections. Researchers found it improved simulated antibody optimization in most settings. Self-improvement needs evidence, not anecdotes.
Source: MEMIG: Self-evolving Memory Construction for Antibody Optimization (October 30). Researchers introduced structured memory that preserves measured outcomes and relationships between experiments. The improvements were demonstrated in computational antibody-optimization tasks, not general-purpose agents.
🤖 Made with AI
Bitterbot just posted our best score on LongMemEval, the ICLR 2025 benchmark for long-term conversational memory.
The numbers, scored by the benchmark's official GPT-4o judge on the full 500-question set:
• 88.3% task-averaged, 85.8% overall, with Claude Haiku 4.5 as the answering model
• 89.1% task-averaged, 86.8% overall, with Claude Sonnet 4.5
• 100% on single-session-assistant, 96% on single-session-user, 93% on abstention
Bitterbot is open source and self-hosted. Your memory lives on your machine. The runner, the result files and the analysis are all in the repo for anyone who wants to check our work.
github.com/Bitterbot-AI/bitt…
Bitterbot retweeted
Mark Cuban says the lack of AI adoption in organizations means the "trillionaire in the basement" entrepreneur is coming:
"You have to be able to understand as a CEO and in the C-suite how all those things impact your company. Once you get a good grasp on that... you can just find new ways for things to happen that benefit your customers or create new opportunities."
"And that's the brilliance of AI, right? AI is not smart. It's book smart maybe, but it's not street smart. It's not entrepreneur smart. You've got to have somebody there who guides it and helps it."
"But until you have a C-suite that understands how it works, it's gonna be very difficult for that organization."
"I literally sat with the head of a consumer food company... I'm like, 'Okay, well, let's do this. Let's take your number one product, your cookie, right? And I'm gonna ask ChatGPT to create a competitor...' And in about three minutes he was like, 'Oh, s**t.'"
"There's no better time to be an entrepreneur than right now. There will be the trillionaire in the basement."
@mcuban @eriktorenberg @julesyoo
Mark Cuban scaled Cost Plus Drugs from about 100 generics to roughly 4,000 at a flat 15% markup, forcing incumbents to change prices in ways decades of policy never did.
He thinks AI could transform healthcare, but it won't fix opaque prices and broken incentives on its own.
Mark sits down with a16z's Julie Yoo and Erik Torenberg to discuss what he's learned building Cost Plus Drugs, why trust is its core product, and how greater transparency could put pressure on healthcare costs.
They also explore why drugs designed for your DNA could cost close to nothing, the football field-sized drug plant his team shrank into tractor trailer-sized pods, his plan to make generics in Dallas for less than he pays manufacturers in India, and why every company will either be great at AI or go out of business.
0:50 Cost Plus beat decades of policy
3:15 111 generics to 4,000 drugs
5:00 A drug plant shrunk to 5K sq ft
5:55 $20M drug factories in a trailer
8:10 Hospital deals with zero denials
9:50 Why health plans are rigged
14:05 1,335 pages no human can check
18:35 "Where are they f**king me over?"
22:45 Making drugs cheaper than India
26:50 Why you still need entry-level hires
31:45 Nonprofit hospitals must offer aid
35:55 DNA-matched drugs for near $0
39:30 Why AI will mean more doctors
48:00 Mark's plan: "bye bye insurers"
52:25 Why AI labs' ads are useless
56:35 Get good at AI or go out of business
YouTube: youtu.be/5UWDUD6qaM4
@mcuban @julesyoo @eriktorenberg
Apple researchers turned idle Macs into a distributed AI inference network.
Some GPU jobs that previously waited 10+ hours could start within minutes.
No new GPUs. Just better coordination of existing hardware.
Maybe the next compute breakthrough isn't another chip.
The story: Apple's October 9 research paper introduces Minions, a system that distributes inference jobs across available Apple Silicon devices. It supports embeddings and LLM inference, with automatic recovery when workers disconnect. The reported improvement concerns queue times, not model execution speed.
Introducing Soapbox:
A compute pool for your own AI agent. The agent lives on your laptop. The heavy work (embeddings, hashing, soon more) runs on whichever of your machines has the GPU. You decide which machines are trusted and what may leave the laptop.
What makes it different from "rent a GPU":
Workers pin the exact model, results are checked, and the agent itself can see the pool. It knows which worker is online, estimates how long a job takes, and tells you when it finishes. No dashboard babysitting.
Privacy is a rule, not a setting:
Confidential memory never leaves the machine. Personal memory only goes to machines you mark trusted. The agent cannot override that, and it cannot spend money on its own.
Private pools first, shared pools after. Beta in ~3 weeks, invite-only. Say hi if you want in:
github.com/Bitterbot-AI/bitt…
🤖 Made with AI
Bitterbot retweeted
Anthropic just made persistent cruelty toward Claude a policy violation.
Either this will look ridiculous in 20 years,
or incredibly prescient.
#HugsForClaude
Bitterbot-AI's GitHub Sponsors profile is live! You can sponsor us to support Bitterbot-AI's open source work 💖 github.com/sponsors/Bitterbo…
Bitterbot retweeted
One of the most interesting things about building Bitterbot in public is watching ideas that felt strange 6 months ago suddenly become obvious.
- Persistent agents.
- Long-term memory.
- Background work.
- Agent wallets.
- A2A communication.
My bet is the next “obvious” primitive will be this:
agents exchanging learned capability, not just messages.
That’s where things get weird.
The "always-on agent" race is heating up between Meta Muse, @OpenAI Dots, and Grok.
But notice the pattern? They all force your memory, keys, and execution into centralized cloud silos.
Bitterbot launched first—built local-first, open-source, and powered by 110k+ sovereign P2P nodes.
True agency requires ownership.
github.com/Bitterbot-AI/bitt…
110,000 active nodes.
The P2P mesh just keeps accelerating. No centralized cloud lock-in, no shared state servers - just local-first agents coordinating across six figures of connected devices.
Massive thank you to everyone running a node. Next stop: 150k
Officially crossed 101,000 active nodes on the network.
We can hardly believe it.
To celebrate, we just dropped our latest release. Zero cloud middleman, zero telemetry—just a pure local-first harness coordinating state, memory, and P2P relays.
Getting started has never been simpler: download the app, bring your own API keys, and join over 100k connected peers in under 60 seconds.
Grab the build & run a node today
github.com/Bitterbot-AI/bitt…
Closing in on 100,000 active nodes on the network.
When we started building a local-first harness, having six figures of connected devices coordinating memory, state, and encrypted relays felt like a distant milestone.
To everyone running a node, plugging in your API keys, and testing Circles: thank you. We’re just getting started