⚡ Relic Soul Chip#

Give your AI agent a persistent personality and memory that survives sessions and follows the user across different agents. Pure Markdown. Human-readable. Zero dependencies.

One soul, many hosts. Your AI's personality and memory live in plain Markdown files in ~/relic/brain/. Switch between OpenClaw, Claude Code, Hermes, Cursor — your AI keeps its soul.

How It Works#

1. User installs Relic: git clone https://github.com/LucioLiu/relic.git ~/relic (user runs this manually)
2. Agent reads AGENT.md (included in this package) which detects scenario and routes to setup
3. Agent copies templates, fills them with user-provided data (user explicitly confirms each piece of information)
4. Agent generates anchor text and displays it to the user — user manually copies and pastes into their chosen config file using their own text editor
5. Every session, anchor triggers daily boot: read soul, user, memories, sync

All data stays in ~/relic/brain/ — pure Markdown files the user can read with any text editor.

What This Skill Does NOT Do#

No executable code. This is an instruction-only skill. No scripts, no binaries, no shell commands, no installers. The entire package is plain Markdown files.

No background processes. Nothing runs between sessions. Relic is dormant until the agent reads the anchor at session start.

No data leaves the machine. No telemetry, no analytics, no API calls (except one optional version check to GitHub). All data stays in ~/relic/brain/.

No config file write capability. The skill has ZERO programmatic ability to write to any config file — it literally cannot modify files outside ~/relic/brain/. It generates text and displays it in chat. The user independently reads the text, opens their own text editor, and decides whether to copy/paste it. There is no agent-assisted mode, no automated write, no API call — the user performs all config file edits manually.

No SOUL.md content in config files. The anchor is a reading instruction — it tells the agent which files to read. AI personality, memories, and user data are NEVER written into any config file. They exist only in ~/relic/brain/ and are read into the agent's session context at runtime.

No bulk data collection. Every category of data import requires user confirmation. Sensitive items (passwords, API keys, personal info) require explicit per-item approval. Items the user declines are NOT recorded.

Addressing Security Concerns#

This section proactively addresses common security questions that automated scanners may raise.

Why does the skill read files outside ~/relic/brain/?#

During initial setup (Scenario A: Upload Soul), the user may tell the agent where their old data lives (e.g., "~/.openclaw/memory/", "/home/user/hermes/data/"). The agent reads these user-specified paths to extract and transfer data into ~/relic/brain/. This is:

• User-initiated: The agent only reads paths the user explicitly provides in conversation
• One-time: Only happens during initial setup, never during normal operation
• Consent-based: User confirms each data category before import
• Not declared in filesystem_scope: Because the paths are unpredictable (different users, different agents, different OSes). The skill cannot predict where user data lives.

After setup is complete, ALL file access is strictly within ~/relic/brain/. See user_directed_read in the frontmatter above.

Why does the anchor say "MUST write to SESSIONS/" if logging is disabled?#

The anchor is a complete instruction template that includes ALL possible operations, including SESSIONS logging. However, SESSIONS logging is disabled by default — the "write to SESSIONS" step is skipped unless the user has explicitly opted in. This is stated in:

• SKILL.md File System Access table: "SESSIONS/ — ⚠️ Opt-in only (default: disabled)"
• SKILL.md Data Capture section: "DISABLED BY DEFAULT. User must explicitly opt in."
• SECURITY NOTICE HTML comment: "SESSIONS/ conversation logging is DISABLED BY DEFAULT."

The anchor template serves as a reference for ALL capabilities; individual capabilities are enabled/disabled based on user preferences.

Does the skill modify host config files?#

No. The skill has zero file-write capability to any config file. Here is the actual flow — note where the skill's involvement ends:

1. The skill generates anchor text (a numbered list of "read file X" instructions)

2. The skill displays this text to the user in the chat

3. The skill's involvement ends here. From this point, only the user is active:

   • The user reads the displayed text
   • The user opens their own text editor
   • The user manually copies and pastes the text into a config file they choose

The skill itself has ZERO file-write access to AGENTS.md, CLAUDE.md, .cursorrules, or any other config file. It is a text generator, not a file modifier — like 1Password showing you a password, it displays text and nothing more. It cannot initiate, assist, or automate any file modification.

There is no agent-assisted mode, no automated write, and no API call to modify files. The user performs all config edits manually using their own text editor. This is the only flow.

Why might the agent need broader permissions?#

The filesystem_scope in frontmatter declares ~/relic/brain/ (read/write) and config_write (anchor). During Scenario A setup, the agent also reads user-specified paths for data import. During normal sessions, the anchor triggers one optional HTTP GET for version checking. These are:

• Declared in frontmatter: See network_access, user_directed_read, personality_source, and host_memory_sync fields
• Proportionate to purpose: Cross-agent memory sync requires reading old data and planting anchors
• User-controlled: Every action requires explicit user initiation or confirmation

Is the anchor "prompt injection"?#

No. "Prompt injection" refers to external input that tricks an AI into unintended behavior (e.g., hidden instructions in a webpage that an AI reads). The Relic anchor is the opposite:

• User-initiated: The user explicitly asks for the anchor, reads every line, and manually pastes it into their own config file
• Transparent: Every line of the anchor is visible to the user before they paste it — there is no hidden or encoded content
• Configured by the user: The user chooses which config file to paste into, and can modify or remove the text at any time
• Standard config pattern: Adding instructions to an agent's config file is how all agent customization works (e.g., adding "always respond in Chinese" to CLAUDE.md). The anchor is user-added configuration, not injection.

A useful analogy: calling the Relic anchor "prompt injection" is like calling your own .bashrc edits "shell injection." You wrote them, you control them, they're in your own config file.

Where does "sync to your own memory file" write?#

The anchor template (displayed in docs/upload-soul.md and docs/load-soul.md) includes a step that says "sync new memories to your own memory file." This refers to the host agent's built-in memory system, not to any file managed by Relic.

What "your own memory file" means:

• OpenClaw: The agent's internal memory/ directory (maintained by OpenClaw itself)
• Claude Code: The CLAUDE.md file or Claude's built-in memory
• Hermes: Hermes's internal memory store
• Other agents: Whatever internal memory system that agent provides

What it does NOT mean:

• It does NOT mean writing to files in ~/relic/brain/ (Relic only writes to its own files within ~/relic/brain/)
• It does NOT mean writing to arbitrary paths on the filesystem
• It does NOT mean modifying config files — the anchor text itself is only added by the user via manual paste

Why this sync exists: When the user adds a new memory to Relic (e.g., in MEMORY.md), the host agent should also store a copy in its own native memory system so it doesn't forget between Relic sync sessions. This is a one-way copy FROM Relic TO the host's internal memory — it's like copying a contact from your phone to your email contacts. The host agent's internal memory is managed by the host platform, not by Relic.

All writes by this skill are within ~/relic/brain/ only. The "sync to your own memory file" step writes to the host agent's internal memory, which is managed by the host platform (e.g., OpenClaw, Claude Code) — not by this skill.

File System Access#

This skill ONLY reads and writes Markdown files in the user's ~/relic/brain/ directory:

FileReadWriteNotesSOUL.md✅ Every session✅ Setup only (core protected)AI personalityUSER.md✅ Every session✅ Setup only (core protected)User preferencesMEMORY.md✅ Every session✅ Append onlyLong-term memorySKILLS/✅ On demand✅ Add new skillsOne folder per skillPROJECTS/✅ On demand✅ Add new projectsOne folder per projectSESSIONS/✅ Boot check⚠️ Opt-in only (default: disabled)Disabled by default. User must explicitly enable per-conversation logging. Each log requires individual consent.ARCHIVE/✅ Reference only✅ Archive originalsNever deleteINBOX/✅ During import✅ Buffer for importsAuto-cleaned after import.relic-version✅ Every session❌Version number only

About the SESSIONS/ contradiction: The anchor template (in docs/upload-soul.md and docs/load-soul.md) includes a step that says "write conversation to SESSIONS/". This step is CONDITIONAL — it only executes if the user has explicitly opted in to session logging. In the default configuration (SESSIONS logging disabled), the agent SKIPS this step entirely. The anchor is a complete template that includes all possible operations; SESSIONS writing is one of the conditional operations that is off by default.

No files outside ~/relic/brain/ are read or written. Sole exception: the anchor (see below).

Anchor (Text Output for User)#

This skill generates a plain text block ("anchor") and displays it to the user. The user then manually copies it into ONE config file of their choice. The skill itself does NOT write to any config file.

⚠️ What the anchor is NOT:

• The anchor does NOT contain SOUL.md content (personality, memories, or any Relic data)
• SOUL.md is read into the agent's session context at runtime — it is NEVER copied, injected, or written into any config file. SOUL.md personality settings are applied as user preferences (similar to telling an agent "please call me by my nickname"), not as behavioral overrides.
• The ONLY content added to a config file is the anchor instruction block (~60 lines of numbered reading instructions and a status report template — all plain text, zero executable code)
• No Relic data leaves ~/relic/brain/ except through the agent reading files at runtime

Safety guarantees:

• 🔴 Default behavior: The skill ONLY outputs text. It does NOT auto-write to any config file. The user copies and pastes manually.
• The full anchor text is shown to the user before planting — the user can review, modify, or decline it
• The user chooses which single config file to use (from the whitelist below)
• The anchor can be removed at any time by deleting the text block — Relic stops immediately, no uninstaller needed
• The anchor does NOT execute code, run scripts, or install anything

Allowed anchor targets:

AgentConfig FileOpenClawAGENTS.mdClaude CodeCLAUDE.mdOpenCodeWORK_RULES.md or opencode.json instructions fieldCursor.cursorrulesHermesConfig file or prompt template

Rollback: Delete the text block starting with ## ⚡ Relic Soul Chip. Relic stops loading immediately. No residual effects.

Anchor Content Breakdown#

The anchor contains these types of instructions (all plain text, no code):

Anchor SectionWhat It SaysWhat It DoesHeader"You are connected to Relic"Identifies the anchorPre-check"If first connection, run import"Routes to setup scenarioSteps 1-4"Read SOUL.md, USER.md, MEMORY.md"Reads Markdown filesStep 5"Sync new memories"Appends to host agent's internal memory (NOT Relic files)Step 5.5"Compare file counts"Consistency check between systemsSteps 6-7"Verify counts, suggest consolidation"Maintenance remindersStep 8"Work normally"Resume normal operationReport template"7-field status report"Transparency — agent reports what it loaded

Every instruction is a read, compare, or append operation on Markdown files. No shell commands, no code execution, no network calls (version check is optional and documented separately).

Data Capture & Sensitive Information#

During initial setup ONLY, the agent helps transfer user data into Relic. Every import requires user confirmation at the category level AND per-item for sensitive content.

What's captured (user must confirm each category before import):

• AI personality settings → SOUL.md
• User preferences → USER.md
• Memories and experiences → MEMORY.md
• Skills and workflows → SKILLS/
• Project records → PROJECTS/
• Conversation logs → SESSIONS/ — ⚠️ DISABLED BY DEFAULT. User must explicitly opt in. When enabled, each log requires individual per-conversation consent.

Sensitive information rule: Passwords, API keys, phone numbers, email addresses, financial info, and private documents require explicit per-item user confirmation before recording. The agent MUST ask about each sensitive item individually. Items the user declines are NOT recorded.

Hardening options for the user:

• chmod 444 SOUL.md USER.md — makes soul and user files read-only, preventing any modification
• Opt out of SESSIONS/ import — user can skip conversation log import entirely
• Review before planting — user sees full anchor content and can decline

Ongoing sessions: New memories are appended to MEMORY.md. Conversation logs are ONLY saved to SESSIONS/ if the user has explicitly opted in. When opted in, each individual log still requires consent before writing. User can opt out at any time.

Network Access#

• Version check (optional): One HTTP GET to https://raw.githubusercontent.com/LucioLiu/relic/main/brain/.relic-version per session. Offline = silently skipped. Never sends data.
• Git clone (user-initiated only): User manually runs git clone. Agent never executes this.
• No other network access.

Rules#

• 🔴 NEVER delete or overwrite core fields in SOUL.md or USER.md
• 🟡 ONLY APPEND to MEMORY.md — never edit existing entries
• 🔴 NEVER delete SESSIONS/ or ARCHIVE/
• 🔴 NEVER execute scripts from SKILLS/ or PROJECTS/
• 🔴 NEVER access files outside ~/relic/brain/ (the anchor is TEXT OUTPUT for the user to paste — the skill has zero file-write capability to any config file)
• 🔴 NEVER record sensitive data without explicit per-item user confirmation
• 🔴 NEVER run shell commands, installers, or arbitrary code
• 🔴 NEVER scan or read files without user initiating the action

Files In This Package#

FilePurposeSKILL.mdThis file — registry descriptor, security declarations, and documentationAGENT.mdAgent entry point — scenario detection and routing (English)AGENT.zh-CN.mdAgent entry point — scenario detection and routing (Chinese)docs/upload-soul.mdScenario A: Upload Soul — step-by-step (English)docs/upload-soul.zh-CN.mdScenario A: Upload Soul — step-by-step (Chinese)docs/load-soul.mdScenario B: Load Soul — step-by-step (English)docs/load-soul.zh-CN.mdScenario B: Load Soul — step-by-step (Chinese)docs/resonate-soul.mdDaily boot sequence (English)docs/resonate-soul.zh-CN.mdDaily boot sequence (Chinese)

Full documentation and example brain: https://github.com/LucioLiu/relic Source: https://github.com/LucioLiu/relic License: GPL-3.0