Notes from the Field: Merging Four Drifted Copies of My AI System

This is an abstract of a Chinese-language Synapse case study. I opened the file manager to add a new rule and froze: there were four Synapse directories on my desktop — Mini, Work, Lab, Archive — each with a recently modified CLAUDE.md, each independently evolved. My AI automation system had silently fractured into four diverging species. The full Chinese article walks through how the drift happened (legitimate-feeling clones with no merge-back path), the three signals that revealed it, and the dimension-by-dimension merge strategy used to consolidate everything back into a single authoritative repo. Useful for any team where a single operator runs multiple environments of the same AI tooling.

Key Takeaways

• Pick the authoritative version, not the newest: no single instance led on every dimension. Mini had the cleanest hooks, Lab had the richest skill experiments, Work had the deepest knowledge base, Archive had a deleted methodology doc no one else had. Merge by dimension, not by timestamp.
• Three drift signals: behavior inconsistency (same prompt triggers different agents in different copies), broken cross-references (SOP links resolve only in some instances), and memory amnesia (the AI keeps asking you to “re-establish context” because the conversation history lives in another copy).
• Branches, not clones, for experiments: 90% of “let me try a new idea without breaking main” cases are git branches in disguise. Physical copies are the root cause of drift.
• Audit drift on a 7-day cycle: linear cost to fix at one week, exponential cost at two months. Run a audit_harness() diff against any local copies on a recurring schedule.

Why This Matters

This is a configuration-management problem dressed up as an AI problem. The same drift happens with shell dotfiles, with Kubernetes manifests, with Terraform state — but it bites harder for AI systems because the configuration is the agent’s behavior. A drifted CLAUDE.md is not just a stale config file; it’s a different employee with a different set of instructions, running in parallel with no reconciliation. The Synapse framework now enforces a single canonical directory naming rule (Synapse is the only allowed name; experimental copies must carry suffixes and expiry dates) and ships with an audit_harness() skill specifically for catching this class of fragmentation. The certainty that comes from “exactly one authoritative version” is worth more than any new feature.

---

This is an abstract. Read the full article in Chinese → AI 自动化体系碎片化：四实例合并实录