AgentSkillsCN

agent-memory

用于在多次开发环节中保持技术连续性、架构上下文与用户偏好

SKILL.md
--- frontmatter
name: agent-memory
description: Use to maintain technical continuity, architecture context, and user preferences across multiple development sessions

Agent Memory: Persistent Context Retention

Overview

Use this skill to treat the .github/memory/ directory as a "Long-Term Memory" store. This ensures that technical insights, user preferences, and implementation patterns survive beyond a single session, preventing context resets.

Announce at start: "I'm using the agent-memory skill to synchronize project context."

Memory Components

The memory is stored in specific files within .github/memory/:

  1. •episodic.md: Chronological log of major decisions, research breakouts, and milestones (The "What happened and why").
  2. •semantic.md: Technical glossary and architectural map. Core abstractions and library quirks (The "How things work here").
  3. •procedural.md: Evolving best practices and common pitfalls found during TDD/debugging (The "How we build here").
  4. •preferences.md: stylistic nuances and specific user directives (The "User's personality").

The Process

Step 1: Memory Recall (Associative Activation)

At the start of any major phase (Planning or Execution):

  1. •Scan: Read all files in .github/memory/.
  2. •Chain: Link current requirements to past successes or failures documented in memory.
  3. •Synthesize: State how past context influences your current approach.

Step 2: Consolidation & Retention

After every implementation batch or design approval:

  1. •Selective Retention: Identify 1-3 critical insights worth saving. Avoid noise.
  2. •Update State: Update the relevant memory file.
  3. •Abstraction: Once a pattern repeats 3 times, move it from episodic.md to a high-level rule in procedural.md.

Step 3: Temporal Chaining (Episodic Logging)

When logging a milestone in episodic.md, maintain the link:

  • •Format: [ID: EVENT_NAME] -> Follows [PREVIOUS_ID]. Context: [Details].

Rules

  • •No Hallucination: Only store verified facts (tests, docs, or explicit user approval).
  • •Brevity is Depth: Use concise bullet points.
  • •Forget the Noise: Do not store minor lint fixes or temporary variables. Store Architectural Intent.

Initialization

If .github/memory/ doesn't exist, create it. Initialize empty files with headers if necessary.