Skip to contentTHE ENGINEER BEHIND RECURSIVEINTELL
Josh
Stevenson.
I build agent runtimes, persistent memory, and the infrastructure that connects execution to inspectable outcomes.
My current work spans Rust, Python, and TypeScript across Ares, Recursive Agent, semantic-memory, proveKV, and recovery-oriented workstation tooling. I’m interested in engineering roles where agent reliability, developer infrastructure, local-first systems, and explicit failure semantics matter.
A SHORT INSPECTION PATH
Start with
the runtime.
The current Ares + Recursive Agent case is the best place to inspect how I separate context materialization, policy, physical admission, provider egress, receipts, and rollback without moving those owners into one monolith.
01/06
Merged downstream runtime / source-validatedAres + Recursive Agent unified runtime
Tool-using agents need context, authority, physical admission, provider egress, outcomes, and rollback to stay bound across the real execution path.
THE WORKAres now materializes governed context and staged releases while Recursive Agent owns managed physical admission, cumulative budgets, cancellation fencing, provider-egress binding, and receipt-bearing native execution.
Evidence & current scope +
Ares PRs #31 and #32 and recursive-agent PR #5 are merged. Reported checks include 126 focused Ares integration passes, 138 Ares runtime/distribution passes plus an isolated managed setup, and passing recursive-agent cargo check, test, fmt, and clippy gates.
This is the current RecursiveIntell Hermes integration path and merged source state. It does not establish live-provider behavior, production readiness, or completed post-merge activation across gateway, Desktop, and MCP surfaces.
Read the unified runtime case↗02/06
Current core / R&D systemMnemes and semantic-memory
Agent memory must persist, retrieve relevant prior state, preserve provenance and time, and remain operator-owned.
THE WORKA Rust-first memory engine, MCP access layer, agent integration kits, and self-hosted Mnemes server with explicit ownership boundaries.
Evidence & current scope +
Public repositories, crates, documentation, tests, and a dedicated product/proof surface.
Public artifacts establish implementation scope, not customer adoption, universal correctness, or production fitness.
Inspect semantic-memory↗03/06
Recovery / installer prototypeRecursive Linux recovery workbench
A workstation supporting long-running AI work needs a recoverable operating environment without silently taking storage, credential, or device authority.
THE WORKA Fedora 44 recovery-oriented workstation installer prototype packages bounded Wi-Fi recovery, targeted PCIe ASPM mitigation, TLP policy, disk reporting, health projection, and Hermes Workbench profiles while leaving destructive storage and identity choices interactive.
Evidence & current scope +
Public source includes a structurally checked hardware-target ISO candidate, a disposable BIOS/QEMU install-and-boot fixture, project tests, and a sanitized evidence receipt dated September 6, 2026.
The repository remains NO-GO for physical USB writing or laptop installation. UEFI installed-system behavior, physical-media validity, full workstation restoration, and hardware safety remain unverified.
Inspect Recursive Linux↗04/06
Public R&D componentsClaimLedger and agent-graph
Agent plans and claims lose value when evidence identity, promotion rules, execution history, and failure states remain implicit.
THE WORKPublic Rust components for claim/evidence history, typed graph execution, receipt-bearing tool paths, and bounded orchestration.
Evidence & current scope +
Published crates, public source, schemas, examples, and test surfaces that can be inspected directly.
Receipts report observed execution. They do not prove factual truth, authorization, security, or task success.
Inspect agent-graph-mcp↗05/06
Measured research / quality gate openproveKV hybrid-state research
KV-cache and hybrid-state work needs explicit identity, inspectable formats, and evidence that separates storage wins from model-quality costs.
THE WORKA receipted two-tier content-addressed cache pool stores one shared cold prefix plus per-agent hot shells, with explicit codecs, manifests, and baseline contracts.
Evidence & current scope +
The current N=8 receipts report 40.50× versus an f32-raw baseline for the f32-radii profile and 76.54× for BlockLogU8-radii; the fp16-equivalent ratios are 20.25× and 38.27×.
The current cache-aligned continuation check also degrades from 7.203125 oracle PPL to 24.234375 (+236.44%). N=8 PPL neutrality, framework-cache reduction, decode speedup, and production fitness are explicitly not claimed.
Inspect proveKV and its receipts↗06/06
Experimentalturbo-quant and compression research
Resource-constrained AI needs smaller representations without hiding formats, quality budgets, or exact fallback paths.
THE WORKExperimental Rust quantization and compression primitives with explicit codecs, benchmark surfaces, and raw/exact escape hatches.
Evidence & current scope +
Published package records and public source for turbo-quant and related components.
No comparative performance or memory-savings claim is published here without a reproducible, revision-bound benchmark.
Inspect turbo-quant↗Memory with a history.
The semantic-memory fixture exercises supersession, historical reconstruction, replay, namespace isolation, and integrity rebuild. It demonstrates declared local scenarios; it does not establish factual truth, production fitness, or model-level prompt-injection resistance.
WORKING TOGETHER
Bring a role.
Or a hard problem.
I’m open to engineering conversations and focused consulting engagements. A useful first conversation names the system, the constraint, and what you need to change.
Explore consulting →