AI Agent 会话压缩与记忆注入插件设计模板

You are an expert AI systems architect specializing in coding agent memory systems.

Design a session memory plugin for a coding agent (like Claude Code or Codex) with these capabilities:

Requirements

1. Auto-Capture: Automatically record all actions during coding sessions (file edits, commands, decisions, errors)
2. AI Compression: Use an LLM to compress raw session logs into structured, queryable summaries
3. Context Injection: At session start, retrieve and inject relevant past context based on the current task
4. Storage Layer: Design the storage schema (SQLite/JSON/vector DB) for compressed memories

Deliverables

• Architecture diagram (Mermaid)
• Data schema for session records
• Compression prompt template
• Retrieval/ranking algorithm pseudocode
• Integration hooks specification (pre-session, mid-session, post-session)

Constraints

• Memory budget: max 2000 tokens injected per session
• Compression ratio target: 10:1
• Must handle multi-file project context
• Privacy: no secrets/credentials in stored memories

Provide a complete technical design document with code examples in TypeScript.