Voltropy

ANNOUNCEMENT: We've built a coding agent that beats Claude Code on long tasks. Today we're releasing it for free. Meet Volt: The coding agent who never forgets. → Dominates Claude Code on the OOLONG long-context benchmark, including at every length from 32K to 1M tokens. → Has unlimited recall. No more amnesia. Volt can code for weeks in a single coherent session. → Massively parallel. One tool call can process thousands of tasks. Like "Map" for LLMs. → Open source and model agnostic. Try Volt today with @openrouter or your API of choice. Volt's performance is the result of a new architecture, Lossless Context Management (LCM), which applies lessons from the history of operating systems and programming languages to LLMs. LCM is like paged virtual memory, except for managing context: - Layer 1. An immutable append-only store of everything that occurs in the coding session. - Layer 2. The active "context window" which functions like a cache layer for navigating to the appropriate section(s) of Layer 1 via a high-fanout DAG. From a user perspective, this feels like an infinite context window, because the model never forgets and performance stays crisp. For the technical details, read our paper: papers.voltropy.com/LCM For the code, go to github.com/voltropy/volt. Or get started with one line: curl -fsSL voltropy.com/install | sh

Lossless Context in OpenClaw feat. Josh Lehman x.com/i/broadcasts/1…

Fun fact about lossless-claw: in addition to solving agent amnesia and enabling infinite-length sessions, it's also very token efficient. Lossless summaries are great for prompt caching. I at about a 90-94% cache hit rate. Thanks to incremental compaction, your context rarely grows beyond 80k tokens before truncating back to 30-40k or less. This means that your model is almost always operating faster, smarter and cheaper since it has less overall context to operate on.

Why doesn't Anthropic just solve compaction by adding LCM to Claude? @Voltropy open-sourced LCM. It's in @OpenClaw. Every frontier lab should use it.

A common line of questions I receive: what does lossless-claw do differently than memory systems? How do the two relate? Should I use both? Here’s the lowdown: Memory systems are good for letting you search for information that’s external to your context window, which are typically “memories” extracted from past/different conversations. This is necessary because: Compaction is lossy: when your conversation gets too big, your agent replaces the whole conversation with a summary. Do this a few times and details from the first conversation are no longer part of the summarized conversation. Your context is split across many sessions: you have conversations with different agents over time and want to be able to reference all of that in your current conversation. Memory systems work okay in the first case and pretty well in the second case. lossless-claw works phenomenally well in the first case and only indirectly addresses the second one. Let’s expand that. Lossless context makes frequent summaries of smaller pieces of context in the background. It keeps your most recent messages around verbatim (the “fresh tail”). As the summaries accumulate, they get combined into summaries of summaries. This lets your agent stay focused: older content is still there, but becomes more “vague” over time — kind of like your own recollection of events. Current messages are always there and never suddenly disappear to be replaced by a summary. This effectively solves the “post-compaction amnesia” problem where your agent seems to suddenly forget important recent details about what you were doing. The reason lossless-claw is called “lossless” though is because your older messages never get truly removed. The incremental summaries replace the messages, but act as “pointers” to them that can be used to expand the source messages back into context. Because the summaries stick around, your agent doesn’t forget about what it can expand should it need to. By contrast, memory systems don’t offer the agent any ideas about what can they can be used to remember. This is why you have to frequently tell your agent to “search its memories” explicitly for something. This feels unnatural and is certainly inefficient. Using lossless-claw means that you can keep one conversation going indefinitely without ever needing to reset. This assesses point (2) from above indirectly: if you don’t need to start new sessions all the time, you don’t need a way to recall information from past sessions! If you work across multiple agents and want to share memories between them, or want to be able to recall information that happened outside of the scope of a conversation (eg meeting notes), you’ll want a memory system. Much of what memory systems are used for is a poor fit for them stemming from overly naive approaches to managing context, which unfortunately are industry-standard. Don’t get me wrong: they’re still useful — I still use one — but they’re not the only tool that agents need to become effective personal assistants. Lossless-claw is among the first production-grade implementations of an alternative context management strategy, and certainly the most effective, and it’s only available on @openclaw. None of this would be possible without the excellent research into Lossless Context Management pioneered by @ClintEhrlich and @rovnys at @Voltropy, so make sure to give them a follow if you’re looking for some real alpha.

"Install Lossless Claw immediately! This is the single biggest upgrade you can make to your OpenClaw setup. Not close." -@DBCrypt0

The Mog programming language. Big in Japan. x.com/matsuu/status/…

Voltropy (W26)

"You won't find a better plugin" - @DBCrypt0 on LCM in @OpenClaw

A story in three parts. LCM "genuinely is going to change my life."

ChatGPT --> OpenClaw --> LCM

"[LCM] does something that no other AI plug-in has ever done before: It gives your AI a permanent memory, that never deletes anything, ever. Your OpenClaw is about to get dramatically more powerful." - @JulianGoldieSEO

I met them at @hf0 can confirm they are crazy

Six weeks ago, we were testing LCM in the basement of @HF0 at 2AM. Today it's endorsed by @steipete as the long-context solution for @OpenClaw. Welcome to the year of Voltropy. It's only March.

The claw father endorses LCM in OpenClaw. 🦞 Voltropy + OpenClaw 🥂

A couple months ago, we invented LCM in a basement. Now millions of people have access to it in @OpenClaw. Feels good, man.

LCM support in OpenClaw is "a big deal." - @chrysb Couldn't agree more.

Support for Lossless Context Management has officially been added to OpenClaw. Congratulations to @jlehman_ on shipping this. Welcome to the world of claws with infinite virtual context. Link below. ⬇️

@davefontenot Risky. Your agent can write code that exceeds its own permissions. Mog fixes that.

@Voltropy how risky is it to run your claw without mog?

ANNOUNCEMENT: We just mogged malware. Introducing Mog, a programming language for self-modifying AI agents. Mog solves the security problems of claws + the usability problems of sandboxing. 1/N 🧵

I'm excited to publish Mog, the world's first programming language designed specifically for extending AI agents. Check this out: an agent could write a Mog script that shells out to bash, *while maintaining the agent's granular bash permissions,* since all Mog I/O has to go through the agent. I'm looking forward to seeing what people do with it. Should be a good plugin language for OpenClaw et. al.