All AI systems

Case study · Gwee Per Ming

Cross-Agent Context Engineering

A persistent second brain that carries evidence-cited context between Claude Code and Codex sessions.

The problem

Coding-agent sessions lose useful context when work moves between tools or restarts. Raw transcript replay is expensive and does not guarantee relevant, verified recall.

My contribution

Built a transcript-native memory pipeline that consolidates evidence-cited, wikilinked memories into Obsidian, then connects Claude Code and Codex through native lifecycle hooks and a hybrid retrieval system.

Architecture and stack

Rust, SQLite FTS5, BM25, vector search, graph reciprocal-rank fusion, all-MiniLM-L6-v2, MCP, CodeGraph, Claude Code and Codex hooks, and a Next.js monitoring console.

  • Rust
  • SQLite FTS5
  • Hybrid retrieval
  • MCP
  • Claude Code
  • Codex
  • Next.js

What runs today

A local-first system integrated into both coding-agent lifecycles. Its console tracks retrieval quality, hook delivery, system condition, and deployment drift.

Outcome

Reached 96.0% Recall@5 and 0.922 MRR across 246,750 turns, while keeping session-start orientations at or below 1,500 tokens.

Proof and source links