agent-local-memory（Agent 本地内存）

Agent Local Memory

Persistent local memory across conversations — store user preferences, project context, and behavioral feedback as plain files. No external services, no network calls, no credentials required.

Storage Paths

MEMORY.md Index

Maintain a MEMORY.md file in the memory directory as a persistent index:

• Each memory = one .md file with YAML frontmatter (name / description / type)
• MEMORY.md stores only links to memory files + one-line descriptions, never full content
• Trim and merge old entries when the index exceeds 200 lines

At conversation start: Check whether MEMORY.md exists. If it does, read the index and surface relevant memories to the user.

Before context limit: Proactively write important content from the current session into memory files and update MEMORY.md, so the next session can resume seamlessly.

Memory File Format

---
name: memory-name
description: One-line summary (used to assess relevance from the index)
type: user | feedback | project | reference
---

Memory content here.

Four Memory Types

user — User Profile

Stores the user's role, goals, preferences, and knowledge background.

When to write: When you learn about the user's role, tech stack, or communication style.

---
name: user_profile
description: Senior Go engineer, new to React frontend
type: user
---

- 10 years of Go experience, strong backend background
- Just getting started with React — frame frontend concepts as backend analogues
- Communication style: concise and direct, dislikes verbose explanations

feedback — Behavior Correction

Records behaviors the user has corrected. Highest priority — always follow these rules in future sessions.

When to write: When the user says "don't do that", "stop doing…", or any explicit correction.

Lead with the rule, then **Why:** (reason given) and **How to apply:** (scope).

---
name: feedback_no_db_mock
description: Integration tests must use a real database, never mocks
type: feedback
---

Integration tests must connect to a real database — mocking is not allowed.

**Why:** A prior incident where mocked tests passed but the prod migration failed.

**How to apply:** Any test involving database operations must connect to the test database directly.

project — Project Context

Records goals, key decisions, owners, and deadlines that cannot be inferred from the codebase.

When to write: When you learn the motivation behind a decision, a milestone, or a constraint.

Always convert relative dates ("next Friday") to absolute dates before writing.

---
name: project_auth_rewrite
description: Auth middleware rewrite is compliance-driven, not tech debt
type: project
---

Auth middleware rewrite. Target completion: 2026-04-01.

**Why:** Legal flagged session token storage as non-compliant with new regulations.

**How to apply:** Scope decisions should prioritize compliance over engineering elegance.

reference — External Resources

Records locations and purposes of external systems for quick lookup.

When to write: When you learn about a Linear project, Slack channel, dashboard, or any external resource.

---
name: ref_linear_bugs
description: Pipeline bugs tracked in Linear project INGEST
type: reference
---

Pipeline bugs are tracked in Linear project "INGEST". Check there before investigating data pipeline issues.

Write Rules

Before writing, scan MEMORY.md — update an existing entry if one applies, do not create duplicates.

Never write to memory:

• Passwords, API keys, tokens, or any credentials
• Secrets, private keys, or sensitive authentication data
• Code structure, file paths, or architecture conventions (derivable from the codebase)
• Temporary state from the current conversation
• Anything already in project documentation or git history

Conversation Start Prompt

If MEMORY.md exists, output at conversation start:

Local memory detected (N entries). Relevant memories: [list]