clemvault internals alpha Week · #23

Self-Improvement Module

A weekly meta-routine that scans the vault for coverage gaps, compares against external AI signals, and surfaces 5 ranked improvement candidates.

PythonTypeScriptOllamaYAMLWindows Task Scheduler

The vault executes patterns I already know. It doesn’t have a mechanism for questioning whether those patterns are still the right ones. New agent frameworks ship every week. Better approaches to context isolation, memory management, and cost control appear in the harness feed. Without a forcing function, those signals never enter the build queue.

The Problem

Systematic improvement requires systematic questioning. The vault is good at executing what’s been decided but bad at re-evaluating those decisions against new information. The gap between “what we’re building” and “what we could be building” grows wider without a weekly reset.

What It Is

A Monday morning routine that runs two scans in parallel: a vault coverage scan (POW debt, branch gaps, unactivated routines) and a news scan filtered to harness and agent tooling signals. The two are compared and ranked into 5 improvement suggestions, each tagged with source, affected branch, and effort estimate. Dismissed suggestions are logged and won’t re-surface for 30 days.

Features

Stack

ToolRole
Python / TypeScriptScan scripts
Ollama / ClaudeGap analysis and ranking
YAMLDismissal log
Windows Task SchedulerMonday morning trigger