memory-builder

Core Principles

• Atomicity: One memory = one atomic conclusion; use links for complexity, not bloated files
• Conclusion-Oriented: Titles must state a result, not a topic (e.g. "Use JWT" vs "Auth System")
• Graph-Centric: Every memory is a node; orphaned memories are failures
• Truth-Only: Capture verified implementation facts and technical decisions

Memory Taxonomy

• Decision: Technical choices and architectural pivots
• Implementation: Specific solutions or state management details
• Learning: Performance findings, optimization results, or post-mortems
• Concept: Core abstractions or domain-specific terminology
• Issue: Problem RCA (Root Cause Analysis) and resolution records

Workflow

1. Extract: Identify core conclusion from conversation or code changes
2. Classify: Assign type; generate English ID and conclusion-oriented Chinese title
3. Link: Search memory/ directory to identify prerequisites or related nodes
4. Persist: Save as .md in memory/ folder using the Chinese title as filename
5. Verify: Ensure single-conclusion focus and valid [[id]] syntax

Guardrails

• No Topic Titles: Reject "Cache Strategy"; force "Cache Home Data for 5 Mins"
• Atomic Enforcement: If a memory covers two conclusions, split it into two linked files
• Silent Maintenance: Automatically update "Leads to" links in parent memories when creating children

Output Structure

---
id: [descriptive-english-id]
type: [decision|implementation|learning|concept|issue]
title: [结论式中文标题]
created: [YYYY-MM-DD]
tags: [relevant, tags, in, english]
---

# [结论式中文标题]

## TLDR
> [最简洁的核心内容概述]

## Connections
- Based on: [[prerequisite-memory-id]]
- Leads to: [[consequent-memory-id]]
- Related: [[related-memory-id]]

## Core Content
[Context, analysis, and final reasoning]

## Artifacts
- `path/to/file.ts` - Implementation reference
- `docs/specs.md` - Related documentation