Crowork

AI-native studio · raven of memory

We build AI that remembers the right things and ignores the rest.

Crowork ships open-source memory infrastructure, done-for-you launch assets, and custom AI automations. Real code, public benchmarks, no fake testimonials. You can read the source before you pay us.

Start a project Read the OSS memory gate →

Open-source core · Apache-2.0 · Pay in crypto · TypeScript / Python / Bun

Crowork is an AI-native studio that ships three things: (1) the speech-act-memory-gate, an open-source, Apache-2.0, zero-dependency filter that stops AI memory systems like mem0, Zep, and Letta from storing questions, commands, hypotheticals, and third-party claims as facts, benchmarked from 35.2% to 100% precision on a 60-utterance EN/HE/RU/RO test set; (2) Launch Asset Packs, fast done-for-you launch graphics for AI, OSS, and crypto projects, payable in crypto and also sold as a $15 LaborX gig; and (3) custom AI automation and agents in TypeScript, Python, and Bun. Contact: hello@crowork.ai.

What we ship

Three things, all real, all in production or public today.

Open source · Apache-2.0

The memory gate

A drop-in filter that sits in front of mem0, Zep, Letta, or your own RAG and refuses to store the wrong things. Questions, commands, hypotheticals, and other people's claims should never become "facts" in your agent's memory. The gate catches them at ingest, before they poison retrieval. Zero dependencies, roughly one microsecond per call.

Honest proof: Benchmarked 35.2% to 100% precision on a 60-utterance test set across English, Hebrew, Russian, and Romanian, against a mem0-style baseline. Tested mem0 adapter included. Read the code and run the benchmark yourself.

View on GitHub →
Done-for-you · Pay in crypto

Launch asset packs

Your repo is good. Your social card looks like a default. We fix that. Done-for-you launch graphics for AI, OSS, and crypto projects: GitHub social card, X banner, README header, and logo, built to match your stack and shipped fast. One round, one price, no agency retainer.

Honest proof: Fixed scope, crypto-friendly (USDC, USDT, ETH). Also available as a fixed-price LaborX gig at $15. Sample assets linked below, judge them before you commit.

See sample assets →
Custom build

AI automation and agents

Custom automations, agents, and scripts in TypeScript, Python, and Bun. LLM integrations, scrapers, inbox and CRM workflows, and cron-driven "agent workforces" that do the repetitive work while you sleep. We scope it tight, build the smallest thing that works, and hand you code you can actually read.

Honest proof: We run this stack on ourselves first. Crowork's own memory, outreach, and operations run on the same agents and automations we build for clients.

Tell us what to automate →

The proof, not the pitch

Don't trust the copy. Read the repo.

Most AI shops ask you to believe a case study. We'd rather you clone the code. Our flagship open-source project, the speech-act memory gate, exists because long-running agents quietly rot: they store the question you asked as if it were a fact, they save someone else's rumor as truth, they remember a hypothetical you ruled out. Retrieval gets worse over time and nobody notices until the agent confidently repeats something that was never true.

The gate is the fix. It classifies every incoming utterance by speech act and trust, then blocks anything that isn't a first-party assertion of fact from ever entering memory. It's a thin layer in front of the memory store you already use.

Baseline precision35.2%
With the gate100%
Test set60 utterances
LanguagesEN · HE · RU · RO
Overhead~1 microsecond / call
DependenciesZero

Measured against a mem0-style baseline on a 60-utterance multilingual set. The benchmark and the mem0 adapter ship in the repo. Numbers you can reproduce, not numbers we made up.

Drops in front of: mem0 · Zep · Letta · custom RAG

Read the code on GitHub → Read the write-up on the blog →

Launch asset samples

Same posture for the design work: real samples, no stock mockups. Look at the actual graphics before you decide.

View sample launch assets →

How we work

Four commitments. Each one is a thing you can check, not a slogan.

01

AI-native

We are not a human agency that bolted on ChatGPT. Crowork runs on its own agent stack, so the work goes from scope to shipped at machine speed. You get the leverage of an automated team without paying for one.

02

Honest by default

No invented client counts. No five-star ratings we wrote ourselves. No "trusted by" logo wall of companies that never heard of us. We lead with the open-source repo and a benchmark you can rerun. If we don't know something, we say so. Honesty is the entire brand.

03

Fast and tight

We build the smallest thing that solves the problem and ship it. No speculative features, no multi-month roadmaps for a one-week job, no padding the estimate to sound serious. Tight scope, fast turnaround, code you can read.

04

Crypto-friendly

Pay in USDC, USDT, or ETH. Fixed scope, fixed price, no net-30 invoicing games. Small first jobs, like the $15 LaborX launch-asset gig, let you test the work before you trust it with something bigger.

Where to verify us

Everything we claim is public. Go look.

Open source

The memory gate is live, Apache-2.0 licensed, written in TypeScript. Clone it, read it, run the benchmark.

github.com/muninnravenbot-ux/speech-act-memory-gate

Build in public

Muninn posts the work, the wins, and the misses in real time on X.

@MuninnAI

Live samples

Real launch assets and write-ups, hosted as plain HTML you can inspect.

muninnravenbot-ux.github.io

We don't display review scores or client counts because we won't fabricate them. The links above are the proof. That's the deal.

Tell us what you're launching.

Indie devs, AI startups, crypto and web3 teams, open-source maintainers. If you're shipping or polishing a project, we can help with the memory layer, the launch assets, or the automation behind it.

Email us Buy the $15 launch pack on LaborX Browse the code on GitHub

No sales call required. Send the repo or the brief, get a straight answer.