local-first, honestly WIP

The Lab

Current experiments in agent reliability: memory, verification, evaluation, and long-session sanity. These are experiments, not products — the WIP labels are load-bearing.

ralto@local ~ % agent --garden

$ sips startup --memory-aware

→ loading session memory, verification hooks, closeout rules… ok

$ openeval run --graders deterministic --trace live

→ grading agent CLI runs locally. accuracy audit queued.

$ storagescope scan ~/

→ scan complete. results stayed on this Mac. nothing uploaded.

$ gnome status

→ current threads: local-first agents · evaluation dashboards · memory-aware tool use · practical AI workflows

Some of it is polished, some of it is duct tape. Follow along on X for the current state of the tangent, or dig into the code on GitHub.

Open threads

what the gnome is tending

🧠

Memory-aware agents

Agents that start sessions knowing what happened last time: memory-aware startup, session closeout, and fresh-context delegation in SIPS.

SIPS on GitHub →

📊

Evaluation dashboards

Deterministic graders, live trace telemetry, and accuracy audits for agent CLIs — figuring out whether the agent actually did the thing.

OpenEval on GitHub →

🔧

Safer tool use

Verification hooks and guardrails so long agent sessions stay careful instead of getting confidently sloppy.

see the hooks →

🌱

Practical workflows

Notes and small workflow details from actually running these tools, written up so other people can skip a few potholes.

read the notes →