Agent Memory

Definition

Agent memory is not a plugin to bolt onto an agent — it is a core function of the agent harness itself. Managing context, deciding what survives compaction, loading system instructions, and presenting metadata to the agent are all invisible decisions made by the harness.

智能体记忆并非附加在智能体上的插件，而是智能体框架本身的一项核心功能。管理上下文、决定在压缩后保留哪些内容、加载系统指令以及向智能体呈现元数据，均为由框架做出的隐性决策。

Details

Why Memory ≠ Plugin (Sarah Wooders / Letta AI)

The common request to “plug a memory system into my agent” fundamentally misunderstands what memory is:

• RAG over past sessions can be a plugin, but retrieval is only a small part of memory
• Even then, simple grep often beats vector search at moderate scale
• The harness makes many invisible decisions that an external plugin cannot control

Claude Code’s Memory Architecture (3-Layer Design)

1. Index layer (always loaded): MEMORY.md — just pointers, ~150 chars/line
2. Topic files (on-demand): detailed knowledge, fetched when relevant
3. Transcripts (never read, only grep’d): raw conversation history

1. **索引层**（常驻）：`MEMORY.md` — 仅包含指针，约 150 字符/行 2. **主题文件**（按需）：详细知识，仅在相关时获取 3. **转录文本**（从不读取，仅通过 grep 检索）：原始对话历史

Key principles:

• Memory = index, not storage — actual knowledge lives outside, fetched on demand
• Strict write discipline — write to file, then update index; never dump content into index
• Staleness is first-class — if memory ≠ reality, memory is wrong
• What they don’t store is the real insight — no debug logs, no code structure, no PR history

核心原则： - **记忆 = 索引，而非存储** — 实际知识存储在外部，按需检索 - **严格的写入纪律** — 先写入文件，再更新索引；绝不将内容直接堆入索引 - **时效性至关重要** — 若记忆与现实不符，则记忆为错 - **不存储的内容才是真正的洞察** — 无调试日志、无代码结构、无 PR 历史

Letta Code’s Memory Architecture

• Git-backed filesystem for agent memory
• Background memory sub-agents that concurrently manage prompt rewriting
• “Context Constitution” — principles for context management

- 基于 Git 的智能体记忆文件系统 - 后台记忆子智能体，并发管理提示词重写 - “上下文构建” — 上下文管理原则

MemGPT Insight

MemGPT was always a stateful agent harness, not a RAG tool. The “memory” emerged from tools the harness exposed for prompt rewriting and external state management.

MemGPT 始终是一个有状态智能体框架，而非 RAG 工具。其“记忆”源于该框架为提示词重写和外部状态管理所提供的工具。

Connections

• Related to: Harness, Skills
• Mentioned in: Memory Is the Harness

相关于：[[ai-agent-architecture/concepts/harness|Harness]]，[[ai-agent-architecture/concepts/skills|技能]] - 提及于：[[ai-agent-architecture/sources/memory-is-the-harness|记忆即 Harness]]