A local harness that keeps models honest.
Flask harness plus Ollama on your hardware. Parallax vector memory recalled every turn. Approval-gated tools. Verified models, failures published.
What Myles is
Myles is a local-first personal AI collaborator. A Flask harness on your machine drives local open models through Ollama, and a bundled Parallax memory service holds semantic vector vaults of your knowledge and conversation history. Relevant memory is injected automatically each turn. By default, nothing leaves the machine. Around that core sit approval-gated tool execution, a say->do guard that refuses to let a model fake completed work, local voice input through on-device whisper, a PWA reachable from your other devices over a private Tailscale mesh, and a DOM-automation bridge with HMAC auth and a hash-chained audit ledger.
How it works
Local execution
A local Flask harness drives open models through Ollama on your own hardware. Conversations, memory, and files stay on disk by default. The default path has no cloud dependency.
Memory vaults
A bundled Parallax service keeps semantic vector vaults of your knowledge and conversation history. Recall is automatic on every turn — no explicit retrieval commands. The vaults are decoupled from the model, so a model swap keeps the accumulated memory intact.
Say->do integrity
The harness detects a model narrating work instead of doing it, and pushes it to act. If the model still cannot act, Myles states plainly that nothing was created or changed, and names the failing model. An honest failure beats a confident fabrication.
Approval-gated autonomy
Four modes — Plan, Ask-first, Let-it-run, Bypass — set how much Myles may do unprompted. Mutating actions pass a real approval gate. The DOM-automation bridge authenticates with HMAC and writes a hash-chained audit ledger.
Why it is different
Myles is not a wrapper around one vendor's API. The harness is model-agnostic, and any local model must earn its place through a 5-level probe suite — single tool call, result use, chained calls, a loaded say->do build task, and finally a real turn through the live harness, judged solely by the artifact it leaves on disk. Sandbox competence alone does not earn the badge; it has been shown to pass models that fail in production. Memory is decoupled from the model, so upgrading the brain keeps the mind. And honesty is enforced in the harness itself: the published verification table keeps its fail rows — two partial, one failed, one without tool support — because we do not publish false claims.
Where it stands today
Myles is a working private beta running daily on the founder's hardware, being hardened toward release. As of July 26, 2026, fifteen local models are VERIFIED through the full probe suite, live-harness gate included, and the report is published with its failures. Recent work, straight from the log: the model picker fix, the honest-failure say->do guard, and the verification agent itself. Rough edges are stated, not hidden — the PWA service worker is currently disabled behind a kill-switch while the mobile experience is reworked. That is the house standard: this status section is trustworthy for the same reason the model table is.
Say hello
If Myles sounds like it belongs in your world, write to hello@memoryforgeai.com — a person reads every note.
Email Myles's people