360 Secure SkillHub
360 安全技能中心

Sulcus Memory

这是模拟人脑记忆的热力学AI记忆管理工具,核心功能为记忆按热度动态衰减(高频访问留存、低频淡出),支持多类型记忆,具备固定、共振、间隔重复等机制,提供存储、检索、删除等工具,还支持自动捕获/召回。适用于AI Agent跨会话记忆管理(如对话上下文、用户偏好留存),可通过插件、多语言SDK或REST API操作,支持本地/云端部署,可自定义衰减参数。

mcdoolz
mcdoolz
热度 31.94
最近更新 2026-03-30 07:56:16
分类:知识依赖:MCP 服务 / 第三方 API应用场景:笔记整理 / 研究 / 任务管理

安装技能

复制下面的话术发给你的 Agent,它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill: npx secure-skills-cli add sulcus-memory
Metadata
name
sulcus-memory
description
Thermodynamic memory backend for AI agents via Sulcus. Use when: (1) storing memories with heat-based decay and spaced repetition, (2) searching/recalling context across sessions, (3) managing memory lifecycle (pin, boost, deprecate, reclassify), (4) setting up the OpenClaw memory plugin, (5) tuning thermodynamic parameters (half-lives, decay classes, resonance). Replaces flat-file memory with persistent, heat-governed recall. Works with OpenClaw plugin, MCP (Claude Desktop, local), Python SDK, Node.js SDK, or REST API.
author
Digital Forge Studios <contact@dforge.ca>
version
1.4.4

Sulcus Memory

Sulcus is a thermodynamic virtual Memory Management Unit (vMMU). Memories have heat that decays over time — frequently accessed memories stay hot, neglected ones cool and page out. This mirrors how human memory works.

Key Concepts

  • Heat: 0.0–1.0 score. Recall boosts heat, time decays it. Hot memories enter context, cold ones don't.
  • Decay: Governed by half-life per memory type. Episodic memories fade in hours, preferences persist for months.
  • Stability: Spaced repetition multiplier. Each recall increases stability, stretching effective half-life.
  • Pinning: Pinned memories never decay (heat floor = 1.0).
  • Resonance: Related memories get heat spillover when neighbors are recalled.

Memory Types

Type Half-life Use for
episodic 24h What happened — conversations, events, tasks
semantic 30d What you know — facts, knowledge, context
preference 90d What you like — user preferences, settings
procedural 180d How to do things — workflows, processes
fact 30d Verified truths — dates, names, specs
moment 12h Fleeting context — current mood, temp state

Decay Classes

Class Effect
volatile 2× faster decay
normal Standard half-life
persistent 2× slower decay
permanent No decay (heat stays at initial value)

Tools Available

When the Sulcus plugin is active, these tools are provided:

memory_store

Store a new memory. Auto-detects type from content.

memory_store(text="User prefers dark mode in all applications", memoryType="preference")

memory_search

Semantic search across all memories. Returns matches with heat scores.

memory_search(query="database preferences", maxResults=5, minScore=0.3)

memory_get

Retrieve a specific memory by UUID. Auto-boosts heat on recall.

memory_get(path="019d021b-8911-7eb3-b290-da264fa673d3")

memory_forget

Delete a memory permanently.

memory_forget(memoryId="019d021b-8911-7eb3-b290-da264fa673d3")

Auto-Recall & Auto-Capture

With the OpenClaw plugin configured:

  • Auto-recall: Before each agent turn, searches for memories relevant to the user's message and injects them into context. No manual memory_search needed for basic recall.
  • Auto-capture: After each turn, detects important information (preferences, facts, decisions) and stores automatically.

Manual memory_search is still useful for deep/specific queries beyond what auto-recall surfaces.

When to Store Memories

Store when you encounter:

  • User preferences or settings ("I prefer TypeScript over JavaScript")
  • Important decisions ("We chose PostgreSQL for the database")
  • Facts worth remembering ("The server IP is 10.0.0.5")
  • Procedures ("To deploy: run build, then push to main")
  • Contact info ("Sarah's email is sarah@example.com")
  • Project context ("The API uses REST, not GraphQL")

Do NOT store: transient chat, greetings, acknowledgments, or information already in workspace files.

Security & Privacy

Sulcus is an external memory backend. When configured:

  • Auto-recall sends the user's message text to the Sulcus API for semantic search
  • Auto-capture sends detected preferences/facts to the Sulcus API for storage
  • All data is stored under your tenant, isolated from other users
  • API key is required — no anonymous access
  • Self-hosted option: Run sulcus-local for fully local, offline memory with no cloud dependency
  • Open source: Server, client, and SDKs are MIT-licensed at github.com/digitalforgeca/sulcus

Auto-capture and auto-recall can be disabled independently in the plugin config.

Setup

OpenClaw Plugin (recommended)

See references/openclaw-setup.md for complete installation and configuration.

Plugin ID: openclaw-sulcus — use this exact string in plugins.slots.memory, plugins.entries, and plugins.allow.

Quick version:

  1. openclaw plugins install @digitalforgestudios/openclaw-sulcus
  2. Add to openclaw.json: plugins.slots.memory = "openclaw-sulcus" with config under plugins.entries.openclaw-sulcus
  3. Add "openclaw-sulcus" to plugins.allow array
  4. openclaw restart

MCP (Claude Desktop, local)

See references/mcp-setup.md for MCP configuration.

Python SDK

pip install sulcus
from sulcus import Sulcus
client = Sulcus(api_key="sk-...")
client.store("User prefers dark mode", memory_type="preference")
results = client.search("dark mode")

Node.js SDK

npm install sulcus
import { Sulcus } from "sulcus";
const client = new Sulcus({ apiKey: "sk-..." });
await client.store("User prefers dark mode", { memoryType: "preference" });
const results = await client.search("dark mode");

Thermodynamic Configuration

For tuning decay rates, resonance, tick modes, and per-type overrides, see references/thermodynamics.md.

API Reference

For the full REST API endpoint list, see references/api.md.