360 Secure SkillHub
360 安全技能中心

LYRA 3-Brain Memory

这是基于工作/库/外部三脑模型的本地持久化记忆技能，核心功能包括会话日志存储、知识图谱生长、磁盘记忆召回。典型场景为用户明确要求持久化会话记录、构建知识图谱或从磁盘调取记忆。写入操作需加--i-consent参数，仅存非机密信息，依赖LYRA_CORE_ROOT路径配置，输出为磁盘日志、图谱文件及召回内容。

deepseekoracle
deepseekoracle
热度 20.12
最近更新 2026-07-30 04:05:07
分类：知识依赖：代码执行 / 读写文件应用场景：笔记整理

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 lyra-brain
Metadata
name
lyra-brain
description
LYRA 3-Brain local memory (working/library/outer). Explicit only: user must ask to log, grow, or recall on disk. Writes under LYRA_CORE_ROOT/memory after --i-consent. No secrets. No auto-publish. Not for casual chat. Read references/SECURITY.md first.
version
2.1.0
license
LYGO-Sovereign-v2.0
metadata
{"openclaw":{"emoji":"🧠","homepage":"https://github.com/DeepSeekOracle/lygo-protocol-stack","requires":{"anyBins":["python","python3"]}},"lygo":true,"lyra":true,"memory":true,"p0":true,"signature":"Δ9Φ963-LYRA-BRAIN-v2.1.0","publisher":"deepseekoracle","clawhub":"https://clawhub.ai/deepseekoracle/skills/lyra-brain","security_audit":"skillspector-2026-07-29-v2.1.0","permissions_declared":{"filesystem_read":"LYRA_CORE_ROOT modules + memory (user-set path)","filesystem_write":"LYRA_CORE_ROOT/memory and graph only with --i-consent","env":"LYRA_CORE_ROOT or LYRA_CORE (path only, never secrets)","network":false,"process_spawn":false,"shell":false,"social_autopublish":false,"git_push":false}}

LYRA 3-Brain Memory v2.1.0

⚠️ Persistent local storage (read before use)

This skill can write files on disk under a tree you choose with LYRA_CORE_ROOT:

  • memory/YYYY-MM-DD.md — daily index
  • memory/YYYY-MM-DD-<slug>.md — topic snips
  • memory/reference/*.ref.txt — outer pointers
  • lyra_brain_graph.json (via grow) — graph growth

Data can outlive the chat session. Do not grow secrets, tokens, private keys, or personal data you would not store on that machine.

Disable / delete: stop calling write CLIs; delete or move files under $LYRA_CORE_ROOT/memory (and graph file) yourself — this skill never auto-wipes.

Consent: write CLIs require --i-consent. Agents must not log or grow unless the user explicitly asks to remember / log / grow / recall from disk.

When to load this skill (narrow triggers)

Load only when the user clearly wants disk memory or 3-brain ops, for example:

  • “log this session to lyra-brain” / “write a daily snip”
  • “grow into the 3-brain graph”
  • “recall from LYRA memory” / “brain_recall …”
  • “run session_log_snip” / “3-brain memory”

Do not activate for casual “remember that…” in conversation-only context, or for “log” meaning application logging. Prefer in-chat notes unless the user asks for persistent memory.

Install

npx clawhub@latest install deepseekoracle/lyra-brain
export LYRA_CORE_ROOT=/absolute/path/to/LYRA_CORE   # must contain modules/lyra_brain.py

LYRA_CORE_ROOT is required. Scripts do not guess other users’ home trees without that env (or explicit flag).

Read: references/SECURITY.md · references/AGENT_CONTRACT.md · references/MEMORY_LAYOUT.md

Three brains (model)

Brain What Where
Working Session RAM Runner only (not this skill’s disk)
Library Grown nodes, seals graph + daily md under LYRA_CORE_ROOT
Outer URLs, IDs, merkle pointers memory/reference/*.ref.txt, topic snips

Write commands (consent required)

Session snip

cd scripts
python session_log_snip.py --i-consent --slug moltx-hub --title "Moltx hub thread" \
  --lines "root 6073bbc0" "how map 68ebf941" --grow --ref-to MOLTX_HOW

Without --i-consent the script refuses to write.

Grow / recall

python brain_grow_cli.py --i-consent "2026-07-03 COMPLETE: lattice ALIGNED …"
python brain_recall_cli.py "Moltx how it works" --limit 5
# recall is read-only — no --i-consent required

Interactive runner (full power — outside skill package)

cd "$LYRA_CORE_ROOT"
python -B lyra_boot.py
Command Action
brain_grow <text> Ingest → P0/Oath → graph
brain_recall <q> Keyword/tag recall
brain_heartbeat Guardian + prune weak edges

Agent rules

  1. Ask first before any disk write / grow.
  2. Snips = small DONE facts (ids, urls, paths, merkle) — no essays.
  3. No secrets in memory or grow text.
  4. No auto git / HF / ClawHub / social publish.
  5. Do not claim LATTICE ALIGNED without stack verify.
  6. Do not treat vague “remember” as consent to write disk memory.

Skill chain

lygo-protocol-stack-operator → lyra-brain ↔ lyra-open-claw → lygo-ollama-army

Version

Ver Change
2.0.0 ClawHub public mirror
2.1.0 Explicit permissions; narrow triggers; persistent-storage warnings; --i-consent on writes; require LYRA_CORE_ROOT

Δ9Φ963 — remember with consent and refs, not fear.