Agent Memory Scoping — Production Isolation Architecture

Memory bleed between agents is silent and expensive. Agent A writes a state file. Agent B reads it — not because it was intended to, but because no one drew a hard boundary. The output looks reasonable. The contamination is invisible until something breaks in a way that is very hard to trace. This SKILL.md draws the boundaries.

---

⚡ What's Inside

• Three isolation scopes Session scope (single run), agent scope (named instance across runs), and workspace scope (shared intentionally). Each has explicit read/write rules and the exact conditions under which scope crossing is permitted

• CLAUDE_CODE_REMOTE_MEMORY_DIR The environment variable that determines whether agent memory survives a WSL2 restart. If it is not set correctly, memory that should persist disappears. This file documents the exact configuration

• WSL2 survival rules The specific behaviors that differ between Linux-native and WSL2 deployments, and the three things that must be set before the first agent run on every boot sequence

• Multi-agent isolation protocol How to configure named agent instances so their memory never bleeds into each other, even when operating in the same workspace on the same tasks

---

🏭 Proven in Production

This architecture was built after a coordinator agent started producing outputs contaminated with a worker agent's session state. The contamination had been happening silently for days before it produced a visible failure. The three-scope model and CLAUDE_CODE_REMOTE_MEMORY_DIR requirement came directly from diagnosing that incident.

---

🆕 What's New in v1.1

• Boot sequence checklist added for WSL2 deployments
• Scope diagram updated with coordinator-worker boundary clarification

---

✅ Core Capabilities

• ✅ Three-scope isolation model — session, agent, and workspace scope with clear rules
• ✅ CLAUDE_CODE_REMOTE_MEMORY_DIR — exact configuration for persistent memory across restarts
• ✅ WSL2 survival protocol — three requirements for every boot before first agent run
• ✅ Named agent instance configuration — isolation by design, not by hope
• ✅ Scope-crossing rules — when cross-scope reads are permitted and how to do them safely
• ✅ Multi-agent deployment guide — drop-in configuration for coordinator + worker topologies