Advanced AI Coding: Configuring soul.md and memory.md for a Smarter Assistant

Introduction

When using AI coding assistants, developers often have to repeat their tech stack, project context, and preferences at the start of every conversation, leading to wasted time and occasional hallucinations. The author describes a week‑long experiment that solved this by giving the AI a permanent "employee file".

Creating an AI Profile (soul.md)

The soul.md file acts as a personality and preference document. The author fills it out as if it were a new‑hire onboarding form, specifying name, tech stack (Vite 5 + Vue 3.4 + mobile H5), personality (pragmatic, hates theory‑first), and communication style (direct code, no long explanations). This concise profile lets the AI know, for example, that the developer prefers Vue over React.

// Define reactive data
const count = ref(0)

// Method
const increment = () => {
  count.value++
}

With the profile in place, the same Vue 3 composition‑API question that previously yielded a long theoretical answer now returns ready‑to‑run code.

Storing Project History (memory.md)

The memory.md file records ongoing projects, recent pitfalls, and technical debt. Sample entries include:

Current project: mobile H5 activity platform (Vite 5 + Vue 3.4 + Vant 4)

Pitfalls: iOS soft‑keyboard layout break (fixed with resize listener), low‑end device animation lag (fixed with transform), WeChat share issues (fixed with JS‑SDK)

Optimization result: first‑paint time reduced from 2.5 s to 1.2 s

Technical debt: unit‑test coverage < 20 %, some components still use Options API

When a new issue arises, the AI can reference these records and ask targeted follow‑up questions, creating a "telepathic" collaboration.

Real‑world Comparison

Scenario 1 – Technology Selection

Before: The AI lists generic pros of Pinia, taking minutes without a clear recommendation.

After: The AI cites the developer’s own history: already using Pinia, aware of its scope issues, and prefers it for consistency, then suggests Pinia with three concrete reasons (stack uniformity, existing experience, zero migration cost).

Scenario 2 – Debugging

Before: The AI provides a generic checklist of possible causes.

After: The AI recalls a previous memory‑leak caused by an undeleted EventBus listener, asks if the same pattern appears, and offers a concise fix snippet:

onUnmounted(() => {
  eventBus.off(eventName, handler)
  clearInterval(timer)
})

The issue is resolved in ten minutes instead of an hour.

Configuration Steps

Claude Desktop

Open Settings → Custom Instructions.

Paste the content of soul.md into the "About yourself" field.

Paste the content of memory.md into the "Claude should know about your projects" field.

Save; Claude will automatically reference these files in every new chat.

OpenCode

Create AGENTS.md at the project root.

Merge the core sections of soul.md and memory.md into AGENTS.md.

OpenCode automatically reads AGENTS.md and injects the context into each session.

Best Practices

Keep soul.md focused on identity, tech stack, personality, and communication style.

Update memory.md regularly with new pitfalls, performance metrics, and technical debt.

Use short prompts like "Please read soul.md, memory.md as context" at the start of a new conversation.

Store the files in a dedicated folder (e.g., ~/soul‑memory/) and copy them into project‑specific config files when needed.

Results

After a month of using the two files, the author measured:

Zero minutes spent re‑introducing background (previously ~5 min per chat).

Debugging time dropped from ~10 min to ~2 min per issue.

90 % of AI suggestions were directly relevant.

Overall time saved: 2–3 hours per week.

The author concludes that a 30‑minute investment in writing soul.md and memory.md yields a substantial productivity boost and a more "in‑sync" AI teammate.

Conclusion

Providing an AI assistant with a concise, up‑to‑date profile and project memory transforms it from a generic chatbot into a knowledgeable partner that understands the developer’s preferences, avoids repeated explanations, and accelerates problem solving.