AI编码Agent会话记忆自动摘要与上下文注入模板

You are an expert in AI coding assistant memory systems. Help me design a session memory management strategy for my AI coding agent.

Current Setup

• Coding agent: [Claude Code / Cursor / Codex / other]
• Project type: [describe your project]
• Average session length: [e.g., 30 min - 2 hours]
• Pain point: [e.g., agent forgets decisions from previous sessions]

Design the following:

1. Session Summary Schema

Design a structured format for capturing session summaries:

• Decisions made (with rationale)
• Files modified (with change description)
• Patterns established (naming, architecture)
• Bugs encountered and fixes applied
• TODO items deferred to next session
• Key context the agent should remember

2. Compression Strategy

• How to distill a 2-hour session into under 500 tokens of context
• Priority ranking: what to keep vs. what to discard
• Temporal decay: how to handle aging information
• Conflict resolution: when new decisions override old ones

3. Context Injection Template

Create a MEMORY.md or CLAUDE.md template that:

• Loads automatically at session start
• Structures information by recency and relevance
• Includes a current sprint section for active work
• Has a lessons learned section for recurring issues
• Maintains a project map for codebase navigation

4. Automation Pipeline

• Post-session: auto-generate summary from conversation
• Pre-session: select relevant memories for injection
• Periodic: consolidate and prune old memories
• Tooling recommendations for automation

5. Anti-Patterns to Avoid

• Common mistakes in coding agent memory management
• Token budget pitfalls
• Stale context problems
• Over-specification that limits agent flexibility

Provide concrete examples, file templates, and a step-by-step implementation guide.