agent-memory — 记忆系统

What This Skill Provides

A complete structured memory system that gives agents persistent, cross-session context. Modeled after Claude Code's memdir architecture but adapted for OpenClaw agents.

Architecture

MEMORY.md          ← 索引文件（最多 200 行 / 25KB），指向 topic 文件
memory/
  user/            ← 用户偏好和习惯
  feedback/        ← 纠偏记录
  project/         ← 项目上下文
  ref/             ← 外部系统引用
  logs/            ← KAIROS append-only daily logs (logs/YYYY/MM/YYYY-MM-DD.md)
  archive/         ← MEMORY.md 溢出归档
memory/heartbeat-state.json  ← 同步状态追踪

Quick Start

First Time Setup

python3 ~/.openclaw/workspace/skills/agent-memory/scripts/init_memory.py

Or manually create the directory structure if you prefer not to run scripts.

Session Startup (read in this order)

1. SOUL.md → who am I
2. USER.md → who is the human
3. MEMORY.md → long-term memory index
4. memory/logs/YYYY/MM/YYYY-MM-DD.md → today's KAIROS log (create if missing)
5. memory/logs/YYYY/MM/YYYY-MM-DD.md → yesterday's log (if exists)
6. memory/heartbeat-state.json → sync state

Then: search memory_search for topic files relevant to current conversation.

When to Write Memories

Memory Writing Rules

Four Memory Types

user/       — Who the human is, their preferences, habits
feedback/   — Corrections, guidance, things to avoid
project/   — Project context not derivable from code
ref/       — External systems (API docs, credentials, services)

Frontmatter Required (every topic file)

---
name: Human-readable title
type: user | feedback | project | ref
tags: [tag1, tag2]
created: YYYY-MM-DD
updated: YYYY-MM-DD
importance: low | medium | high
---

MEMORY.md Index Rules

• Update MEMORY.md when adding/removing topic files
• Do NOT update MEMORY.md when just editing topic file content (update the file's updated field instead)
• Hard limit: 200 lines AND 25KB — whichever hits first triggers archiving
• Archive low-importance entries to memory/archive/ when approaching limits

KAIROS Daily Log (append-only)

Path: memory/logs/YYYY/MM/YYYY-MM-DD.md

Format:

# YYYY-MM-DD

## HH:MM - Event Title
- Description
- Key decision
- Follow-up action

Rule: Append only. Never edit past entries. Create date-named files.

Maintenance (Heartbeat Integration)

Every heartbeat, check memory/heartbeat-state.json:

import json
state = json.load(open("memory/heartbeat-state.json"))
if state.get("lastMemorySync") is None or \
   hours_since(state["lastMemorySync"]) > 4:
    # Run: python scripts/memory_sync.py

Nightly distillation (02:00-03:00 local):

if is_between("02:00", "03:00") and \
   hours_since(state.get("lastDistillation", 0)) > 12:
    # Run: python scripts/distill.py

Workspace scan (every 3rd sync):

if state.get("memorySyncCount", 0) % 3 == 0 and \
   hours_since(state.get("lastWorkspaceScan", 0)) > 8:
    # Scan workspace for new/modified files worth remembering

Key Patterns

Proactive Retrieval (before answering)

Evan: "继续上次那个项目"
→ memory_search("上次 项目")
→ Found: memory/project/my-project.md
→ Load context, then proceed

Writing a New Memory

Evan: "我喜欢用 zsh，不用 bash"
→ Determine type: user
→ Write to: memory/user/shell-preference.md (with frontmatter)
→ Update MEMORY.md index
→ Log to today's KAIROS: "Evan 偏好 zsh"

Tool Selection (using TOOLS.md)

Before using exec: consult TOOLS.md for risk_level and permission
Before searching: consult TOOLS.md for search tool options

References

• Memory types: See references/memory-types.md for detailed type descriptions
• Frontmatter: See references/frontmatter.md for schema spec
• Startup workflow: See references/startup-workflow.md for complete procedure

Scripts