Field notes on compaction drift and repo memory.

Compaction drift is the new bottleneck

Long agent sessions compact away the reasoning that keeps the plan stable. This post unpacks why drift happens and how to preserve intent when the window tightens.

Local-first repo memory vs cloud vectors

Cloud vector stacks optimize for hosted apps; local-first wins for repo context. We break down privacy, cost, and latency tradeoffs.

Designing MCP tools for bounded context

Bounded tool outputs keep the plan alive without sacrificing evidence. Here's how to design MCP tools that stay useful under a tight token budget.

Symbol graphs beat raw tokens

AST and reference graphs expose relationships that text search misses. This is why symbol graphs are essential for refactors and investigations.

Context Bandwidth Credits: pricing the pain

Meter context delivery instead of storage to align cost with value. CBC pricing ties spend to the thing that actually hurts: the window.

Compaction recovery playbooks

A repeatable reset protocol for long agent sessions. Recovery playbooks turn compaction into a quick routine instead of a restart.

Built for regulated teams

Local-only indexing and controlled deployments unlock adoption in regulated environments. Security posture is a product feature, not an afterthought.

Measuring plan stability

Turn compaction-drift intuition into trackable metrics. Measure plan stability and context efficiency over time.