1. X
  2. Josh Stevenson | RecursiveIntell
Log inSign up
Josh Stevenson | RecursiveIntell
9,018 posts
Image
user avatar
Josh Stevenson | RecursiveIntell
@RecursiveIntell
Rust AI systems engineer | Building semantic memory, MCP infrastructure, edge AI & compression. Open source: RecursiveIntell.com
Albertville, AL
recursiveintell.com
Born August 1, 1984
Joined February 2025
3,376
Following
1,461
Followers
RepliesRepliesArticlesArticlesMediaMedia
  • user avatar
    Josh Stevenson | RecursiveIntell
    @RecursiveIntell
    6h
    proveKV multi-agent state machine just ran live. 5 LLM calls, 6 nodes, 3 parallel agents, 9.1 seconds. Cryptographic receipts for every state transition. What this proves: the proveKV executor seam is wired end-to-end. Every Agent Graph node now receives a proveKV handle
    Image
    GitHub - RecursiveIntell/proveKV: Two-tier, receipted, content-addressed KV-cache pool. fib_k4_n32...
    From github.com
  • user avatar
    Josh Stevenson | RecursiveIntell
    @RecursiveIntell
    12h
    This is the kind of discipline i've instilled in my agents. It will not let the slightest issue slide. This is me implementing a VERY difficult plan for my recursive agent. GPT-5.6 Sol Max has been struggling, 5+ attempts and still hasn't got it right. @OpenAIDevs
    Image
  • user avatar
    Josh Stevenson | RecursiveIntell
    @RecursiveIntell
    13h
    I may have just changed how we setup and think about agents. Now an agent on one devices has access to the actual kv cache of another running on a separate device. It can now be thought of as a variable you use. @NousResearch (haven't tested yet, but should work)
    Image
  • user avatar
    Josh Stevenson | RecursiveIntell
    @RecursiveIntell
    16h
    check this out...it's so efficient now, an arduino uno q 4GB can run 12-16 Hermes agents at once with headroom left. If that edge device can run that many now, the possibilities are endless. @NousResearch @Teknium
    Image
  • user avatar
    Josh Stevenson | RecursiveIntell
    @RecursiveIntell
    17h
    I may have solved the cost of KV cache with my two tiered codec, fib-quant cold pool, turbo-quant hot shell, on a poly-kv substrate. You combine this with the memory efficiency from removing langchain/langgraph from Hermes and it feels like a local AI revolution is coming.
    Image

Log in or sign up for X

See what’s happening and join the conversation

Continue with phone
or
Log in with username or email
Terms·Privacy·Cookies·Accessibility·Ads Info·© 2026 X Corp.
Advertisement
Advertisement