Kleos Research

Building the layer between you and your AI Agent.

[Kleos Research @Copilot0x]

Kaleidoscope — local memory for coding agents

Your agent forgets everything between sessions: same questions, same rediscovery, same mistakes. Kaleidoscope gives it a memory that lives in a vault on your own machine. Available macOS or Linux, Node 18+.

Paste this to your agent:
"npm install @kleos-research/kaleidoscope globally, activate it with
this key <key>, then run kscope init here."

That wires up your agent — Claude Code, Codex, Cursor or OpenCode. Restart it after, then work normally: ask it to search your memory, or tell it to remember something.

Kleos Research - https://t.co/deq40ObZH3
Kaleidosocpe - https://t.co/z1bi25Gm0F

If you want to be a alpha tester please DM, I will share API keys

[Kleos Research @Copilot0x]

We help you stop burning tokens and let your Luna perform at Mythos level

[Robinhood Alpha @RobinhoodAlphas]

Kleos Research is building the layer between you and your AI Agent.

As AI agents become more capable, the real challenge is making their intelligence accessible, useful, and easy to interact with. Kleos Research is positioning itself right in that layer — connecting users with the growing world of autonomous AI.

This is more than just another AI narrative. It is focused on the interaction between humans and their agents, where the next generation of AI experiences will be built.

They launched CoPilot and are now launching Kaleidoscope memory layer for agents

C.A: 0x0a1fCBAc65270db349cb8c1EcF05810AC28ED9e8

App: https://t.co/anHD4R9pZ0

Website: https://t.co/09Z0nWii1A

$CPILOT is one to watch as the AI agent narrative continues to evolve.

[Robinhood Alpha @RobinhoodAlphas]

Kleos Research is building the layer between you and your AI Agent.

As AI agents become more capable, the real challenge is making their intelligence accessible, useful, and easy to interact with. Kleos Research is positioning itself right in that layer — connecting users with the growing world of autonomous AI.

This is more than just another AI narrative. It is focused on the interaction between humans and their agents, where the next generation of AI experiences will be built.

They launched CoPilot and are now launching Kaleidoscope memory layer for agents

C.A: 0x0a1fCBAc65270db349cb8c1EcF05810AC28ED9e8

App: https://t.co/anHD4R9pZ0

Website: https://t.co/09Z0nWii1A

$CPILOT is one to watch as the AI agent narrative continues to evolve.

[Kleos Research @Copilot0x]

@virtuals_io @virtuals_intern @buildonvirtuals @aixbt_agent

[Kleos Research @Copilot0x]

Kaleidoscope — local memory for coding agents

Your agent forgets everything between sessions: same questions, same rediscovery, same mistakes. Kaleidoscope gives it a memory that lives in a vault on your own machine. Available macOS or Linux, Node 18+.

Paste this to your agent:
"npm install @kleos-research/kaleidoscope globally, activate it with
this key <key>, then run kscope init here."

That wires up your agent — Claude Code, Codex, Cursor or OpenCode. Restart it after, then work normally: ask it to search your memory, or tell it to remember something.

Kleos Research - https://t.co/deq40ObZH3
Kaleidosocpe - https://t.co/z1bi25Gm0F

If you want to be a alpha tester please DM, I will share API keys

[Kleos Research @Copilot0x]

@composio shall we collaborate? We can add you to our beta testing group.

[Kleos Research @Copilot0x]

1/ We ran Kaleidoscope against BEAM — the long-term-memory benchmark — at both the 100K and 1M scale. 990 questions, 74,658 memories in the 1M tier.

The most useful result is one we weren't looking for: retrieving more memories doesn't produce better answers. 🧵

[Kleos Research @Copilot0x]

7/ Everything is public — per-question rows for every run

https://t.co/5GuG1ilYxc

[Kleos Research @Copilot0x]

@buildonvirtuals @virtuals_io

[Kleos Research @Copilot0x]

Virtual's team PR https://t.co/IGLu0DF7ES
@buildonvirtuals
@virtuals_io

[Kleos Research @Copilot0x]

https://t.co/wODRzNrGqv

[Kleos Research @Copilot0x]

A better model was available at the same size. We shipped the smaller one.

The benchmark said it was 20% worse. Measured where it actually runs: 4%.

Ten encoders, five corpora, fifty quantisation configs. https://t.co/IS2Bu9Uoes

[Kleos Research @Copilot0x]

What if @mem0ai was free if you held $CPilot?

We will be publishing our benchmark for Kaleidoscope on BEAM this week. Kaleidoscope will be free for $CPilot holders.

https://t.co/cpMHuWqF1i

[Kleos Research @Copilot0x]

0xCopilot optimizes for cost per task &amp; completion correctness.

We will publishing benchmarks soon.

[Composio @composio]

We ran DeepSeek V4 Flash through 4 agent harnesses (Claude Code, Codex, OpenCode, Oh My Pi) on 30 agentic tasks.

A different harness won on each metric: success rate, cost, and speed. 🧵🧵🧵

[EtherMage @ethermage]

Thank god for deepseek v4 flash,
when u are building products for adoption, cost is actually a massive factor vs using it for urself to code.

V4 flash from our internal testing is smart enough to do stuff and is 2 orders of magnitude cheaper for the end user. Makes a lot of use case viable now

More stuff for u guys to play with really soon…