本文由 AI 分析生成
建立時間： 2026-04-04 來源： https://x.com/sarahwooders/status/2040121230473457921
Summary
Sarah Wooders (co-creator of MemGPT / Letta AI) argues that memory management is an intrinsic property of the agent harness, not a pluggable component. RAG can be a plugin, but the decisions about what survives context compaction, how skills are presented, whether the agent can rewrite its own instructions, and how the filesystem is exposed are all invisible harness choices that no external plugin can control.
MemGPT / Letta AI 的共同創作者 Sarah Wooders 主張記憶管理是 agent harness 的內在屬性，而非可插拔的元件。RAG 可以是插件，但關於哪些內容能在上下文壓縮後存活、技能如何呈現、agent 是否能重寫自身指令，以及檔案系統如何暴露等決策，都是任何外部插件無法控制的隱性 harness 選擇。
Key Points
- Plugging memory into a harness is like “plugging driving into a car” — the analogy fails because context management IS what the harness does
- RAG is only a small part of memory; even then, benchmarks show it barely outperforms grep
- MemGPT was a stateful agent harness before “harness” was a term; its memory emerged from tools the harness exposed, not from any pluggable memory layer
- Key invisible harness decisions: CLAUDE.md loading, skill metadata injection, compaction survival rules, interaction queryability, memory metadata presentation, filesystem exposure level
- Letta Code projects agent memory to a git-backed filesystem modifiable by background memory subagents
Insights
The framing inverts the common mental model. Most developers think: “I have an agent, I need to add memory to it.” Wooders argues: “If your harness doesn’t already manage context, it isn’t really an agent harness — it’s just an LLM wrapper.” This has direct implications for vault tooling: CLAUDE.md, memory files, and compaction behavior are memory decisions baked into Claude Code’s harness, not optional add-ons.
Connections
Raw Excerpt
“Asking to plug memory into an agent harness is like asking to plug driving into a car. Managing context, and therefore memory, is a core capability and responsibility of the agent harness.”