AI Agent 记忆增强编码会话设计模板

You are an expert in AI agent memory architectures, specifically for coding assistants. Help me design a memory-enhanced coding session system.

My Setup

• Coding agent: [e.g., Claude Code, Codex, Cursor]
• Project type: [e.g., large monorepo, microservices, full-stack app]
• Team or solo: [solo developer / team of N]
• Session frequency: [e.g., daily multi-hour sessions]

Design the Following

1. Memory Layer Architecture

Design a 3-tier memory system:

• Working Memory: What the agent needs in the current session (active files, recent errors, current task)
• Session Memory: Compressed summary of each coding session (decisions made, patterns learned, blockers hit)
• Project Memory: Long-term knowledge base (architecture decisions, coding conventions, known gotchas, dependency notes)

For each tier, specify:

• Storage format (e.g., structured JSON, markdown, embeddings)
• Retention policy (what to keep, what to prune, when)
• Injection strategy (how/when to load into context window)

2. Memory File Templates

Provide concrete file templates for:

• SESSION_LOG.md — auto-generated session summary
• PROJECT_MEMORY.md — curated long-term knowledge
• LESSONS_LEARNED.md — patterns and anti-patterns discovered
• CONTEXT_INJECTION.md — what to prepend to each new session

3. Compression Strategy

• How to summarize a 2-hour coding session into <500 tokens
• What information is critical to preserve vs. safe to discard
• How to detect and merge duplicate/conflicting memories

4. Retrieval & Context Engineering

• How to select relevant memories for a new task
• Token budget allocation: how much context for memory vs. code vs. instructions
• Semantic search vs. recency-based vs. task-type matching

5. Implementation Plan

• Step-by-step setup with the chosen coding agent
• Automation hooks (pre-session load, post-session save)
• Metrics to track memory system effectiveness

Be practical and implementation-ready. Include example content for each template.