Skip to main content
recursiveintelljosh stevenson · the dojo
online · shipping dailysummon roko ↯

Now

Current Focus

Rust-first local AI runtimes, evidence-aware memory, and autonomous agents.

The primary focus right now is the Rust evidence-runtime stack, autonomous AI, and embedded edge-AI: ESP32-S3 Sentinel, AiDENs closed-loop self-learning, semantic-memory as a published MCP server, and the broader RecursiveIntell library workspace.

ESP32-S3 Sentinel

A $4 always-on ESP32-S3 sentinel that runs a local char-LSTM decision tier and only wakes a richer gateway when confidence drops. 11.6 tok/s TinyStories H512 on real Freenove WROOM N8R8 hardware.

  • 6.34M-parameter char-LSTM, hidden=512, 3 layers
  • 4.8 MB int8 weights + int4 recurrent weight file
  • Tier 0 sentinel + optional Tier 1 Ollama gateway over WiFi
  • Worked example runs on a laptop with no hardware required
  • Receipt-backed: every decision logged as JSONL

View the project page or open the repo.

AiDENs Autonomous Loop

AiDENs is a closed-loop self-learning AI system built in Rust. It audits its own typed knowledge graph to find structural and content-level gaps, generates prioritized tasks to fill them, executes those tasks via a local LLM (Ollama), captures results as provenance-attributed facts with graph edges, evaluates fact quality through a governance gate, and records RL routing feedback for adaptive retrieval.

  • 14 iterations demonstrated, 12/12 tasks completed, 29 facts captured.
  • 8 mission types with adaptive priority scheduling: verify published crates, detect contradictions, verify file references, verify codebase sync, trace provenance chains, stale date detection, find duplicates, and audit namespace completeness.
  • 56K LOC across 36 crates with 668 tests (506 core + 162 autonomous).
  • In development — not yet published to crates.io.

semantic-memory

The knowledge graph substrate that powers AiDENs is now published on crates.io. It provides typed edges (semantic, temporal, causal, entity), bitemporal search, contradiction detection, factor graph belief propagation, and RL-trained adaptive retrieval routing.

  • 48 MCP tools with tool profile gating (lean / standard / full)
  • 15 HTTP endpoints for programmatic access
  • RL routing with persistence across sessions
  • Auto-management: integrity checks, vacuum, re-embed via HTTP
  • Available on crates.io: semantic-memory and semantic-memory-mcp

semantic-memory-claude-kit

A Claude Code plugin for semantic-memory integration. Version 0.5.2 is available on GitHub with auto-recall, auto-capture, dedup guard, receipts, primer, and maintenance hooks.

Published Crates

The following crates are published on crates.io:

  • semantic-memory — typed knowledge graph with provenance
  • semantic-memory-mcp — MCP server binary
  • turbo-quant — quantization research (4K+ downloads)
  • claim-ledger — claim and evidence compiler
  • stack-ids — shared identifiers and trace primitives

Other Active Tracks

  • Recall — enforcing daemon-owned authority, runtime truth, receipts, doctor reports, and repair packet generation.
  • Gloss — proving desktop chat produces visible tokens, visible errors, or durable attempt traces for every prompt.
  • ClaimLedger — turning raw claim compilation into source-spanned bundles, support judgments, contradiction records, review queues, and testimony exports for Gloss.
  • Libraries — keeping canonical Rust crates, satellite utilities, and quantization research separated by explicit support boundaries.

Last updated: July 2026

Want more detail? Browse the Projects, the Lab, or the latest Writing.