Datasets:

LucioLiu
/

relic

Relic is a zero-dependency, pure-file, human-and-AI-readable AI personality persistence framework. Any AI Agent can connect by reading the protocol once — no code required.

Core Philosophy

Relic chip = Soul + Memory + Skills (central brain)
Agent (Host) = Execution + Tools (replaceable)

Relic doesn't depend on any specific Agent framework
Agents are replaceable hosts, Relic is the one soul

You (the user) have one soul chip (Relic) storing your personality, memories, and preferences. You can plug this chip into any agent host (OpenClaw, Hermes, Claude Code, OpenCode...). The host gains a soul and works for you. Change hosts, keep the soul.

Design Principles

1. Convention over Configuration

File names are the convention: SOUL.md = soul, USER.md = user, MEMORY.md = memory.

2. Native Markdown

Humans edit directly, agents read directly. cat SOUL.md shows the soul.

3. Three-tier Permission (Anti-Cyberpsychosis)

Permission	Files	What agents can do
🟡 Enrichable (core protected)	SOUL.md, USER.md	May add content but must NOT modify core fields (name, mission, user's preferred name)
🟡 Archive Write	ARCHIVE/	Written only via consolidation (never delete existing)
🟡 Append Write	SESSIONS/	May create new files, must NOT modify/delete existing
🟡 Append Only	MEMORY.md	Only append to the end, never modify old content
🟢 Read/Write	SKILLS/, PROJECTS/	May add and update

4. Append-only Memory

Never modify history, only append. Periodic consolidation archives old entries; originals are never deleted.

5. Progressive Loading (6 layers)

L0 Soul (loaded every session, ~1K tokens): SOUL.md + USER.md
L1 Memory (loaded at startup, ~2-8K tokens): MEMORY.md full file
L2 Skills (loaded on demand): Relevant skills from SKILLS/
L3 Projects (loaded on demand): Relevant project docs from PROJECTS/
L4 Archive (loaded on demand): Consolidated memories from ARCHIVE/
L5+ Sessions (never auto-loaded): Full conversation logs in SESSIONS/

6. Neural Anchor in Agent (Critical!)

Plant an anchor in the agent's own core config file, ensuring the agent remembers to read Relic every session. This is the prerequisite for the entire system — if the agent forgets Relic, nothing works.

7. Bilingual (Chinese + English)

File names in English (easy for agents to reference), content bilingual (everyone can read). UI supports language switching.

Core File Structure

brain/
├── SOUL.md                  # 🟡 Enrichable (core protected) — Soul (includes Core Mission)
├── USER.md                  # 🟡 Enrichable (core protected) — User profile
├── MEMORY.md                # 🟡 Append Only — Refined memories
├── protocol.zh-CN.md         # Agent protocol (Chinese)
├── protocol.md               # Agent protocol (English)
├── .relic-version           # Version number
├── SKILLS/                  # 🟢 Read/Write — Skills library
│   └── *.md
├── PROJECTS/                # 🟢 Read/Write — Project docs (agent or user added)
│   └── *.md
├── SESSIONS/                # 🟡 Append Write — Raw conversation logs (must NOT modify/delete existing)
│   ├── 2026-04/
│   │   ├── 2026-04-15.openclaw.md
│   │   └── 2026-04-16.opencode.md
│   └── ...
├── ARCHIVE/                 # 🟡 Archive Write — Consolidated archives, written only via consolidation (never deleted)
│   ├── 2026-04.raw.md
│   └── index.md
└── INBOX/                   # 🟢 Read/Write — Import buffer

Three Integration Scenarios

Scenario A: Import (Empty Relic + Agent with Memory)

Full house-moving — not just memories, skills, conversations, and projects all come along.

Inventory — Agent scans all persistent content, lists for user confirmation
Import memories → MEMORY.md (organized by category, sensitive info asked proactively)
Import skills → SKILLS/ (no source restrictions)
Import conversations → SESSIONS/ (original content unchanged, unsummarized)
Import projects → PROJECTS/ (identified from memories and conversations)
Cleanup and Confirmation — Fidelity check (original vs converted bytes, warn if >50% compression), archive originals to ARCHIVE/raw/ (never delete), report import statistics

Scenario B: Inject (Relic with Content + Empty Agent)

B-0. Self-check: Assess agent storage capacity, choose transport strategy B-1. Execute boot sequence (read SOUL/USER/MEMORY, sync new memories to self) B-2. Soul injection: Write SOUL.md into agent core config (priority: core identity > behavioral norms > detailed settings) B-3. Sync Relic content (memories/skills/projects, check each item) B-4. Deferred Writes: Do not modify any Relic files during first injection, stage writes until user confirms B-5. Plant Anchor: Embed Relic anchor in agent core config, ensure auto-load on next session B-6. Verification and Report: Coverage check + full sync report to user

Empty Relic + Empty Agent: Ask user — [A] Start fresh: describe yourself, agent fills SOUL.md [B] Use agent defaults: import agent's built-in persona

Scenario C: Merge (Both Have Content)

Three choices, no complex overlap calculation: [1] Keep Relic soul (import agent memory only) [2] Keep agent soul (replace Relic soul) [3] Merge — agent generates preview report, lists key differences, user decides item by item

Memory Consolidation

Dual-line system:

Soft limit 8K tokens: Warn when exceeded, allow continuing
Hard limit 16K tokens: Must consolidate when exceeded

Consolidation process:

Move entire MEMORY.md to ARCHIVE/YYYY-MM.raw.md (original preserved)
Agent generates compact version of MEMORY.md
Update ARCHIVE/index.md

SESSIONS/ (raw conversations) are never consolidated, never auto-loaded. User can choose retention period (forever/1yr/6mo/3mo), default forever. Space isn't an issue — one year of conversations ≈ 72MB, smaller than ten photos.

Single-file Export

When exporting a .relic file (zip package), three-tier distillation:

Level	Contents	Best for
Soul Mode	SOUL.md + USER.md + MEMORY.md	Public sharing
Standard Mode	Soul + Skills + Projects	Device migration
Full Mode	Includes all SESSIONS/	Complete backup (⚠️ privacy warning)

Neural Anchor Guide

After connecting to Relic, an anchor must be planted in the agent's own core config file. Full anchor template in docs/protocol.md Section 9.

Anchor locations per agent:

OpenClaw → AGENTS.md
Claude Code → CLAUDE.md
Hermes → config file or prompt template
OpenCode → WORK_RULES.md or instructions

v1.1.1 Changes

Based on real OpenClaw testing of v1.0.4, 27 issues were identified and addressed:

Concept Shift

From "insert=gain soul, remove=factory reset" → "central brain + bidirectional sync"
Every agent is both reader AND writer — grow-only (G-Set)
Anchors changed from "optional" to "strongly recommended" with consequences explained

PROTOCOL Restructure

PROTOCOL.md: 864 → 475 lines (streamlined, no information loss)
PROTOCOL.en.md: 823 → 475 lines (mirrors Chinese version)
SESSIONS description: "never auto-loaded" → "must create and write, don't auto-read on startup"

README Rewrite

Complete rewrite as beginner-friendly decision tree
No unexplained jargon in first 30 lines
Competitor comparison table removed
chmod explained in plain language
All 6 file types described (not just 3)

Permission Label Fix

SOUL.md and USER.md unified to 🟡 (Enrichable, core protected) across all docs
INBOX/README "delete" contradiction fixed → moved to ARCHIVE/raw/ permanently

Bilingual Coverage

All non-memory files now have bilingual versions
~15 new/updated files for complete Chinese+English coverage

example-brain Expansion

Added SKILLS/, PROJECTS/, SESSIONS/, INBOX/ with Silverhand-style examples
Anti-idiot notes in all directory READMEs

Version Unification

All version strings unified to 1.3.0 (1.1.1→1.2.0 history in older paragraphs; 1.3.0 adds alignment check and structured boot report)

v1.3.1 (2026-04-26)

Protocol Consistency Fixes

SESSIONS/ permission changed from "Read Only" to "Append-Write" across all docs
load-soul B-1 split into read-only phase (Steps 0-4) and subsequent write phase
resonate-soul: added Step 7 (consolidation check), step numbering now matches protocol
PLAN.md: updated 13 outdated design assumptions
Fixed protocol.zh-CN.md quick-reference card table rendering
Translated 4 integration guides and TROUBLESHOOTING.md to English

v1.3.2

ARCHIVE/ changed from Read-Only to Archive-Write semantics (🟡 Archive Write)
MEMORY.md append-only rule now includes header metadata exception
Fixed anchor template Step 4/5 overlap (Step 4 keeps >400 safety gate, Step 5 deduplicates, Step 7 handles >200 suggestions)
Added B-4 Deferred Writes step, B-step renumbered (B-4→B-5, B-5→B-6)
Scenario A: added SOUL/USER import steps + anchor planting step
Scenario C expanded to C-0 through C-5 complete workflow
Integration guides: unified anchor format (summary + pointer), added Step 0/5.5
Agent Registry changed from HTML comments to visible code blocks
resonate-soul: added read/write phase separator
SKILL.md write scope expanded to include SKILLS/, PROJECTS/, INBOX/

Discussion History

⚠️ The following describes early design discussions. Some content has been superseded by the current protocol. Refer to docs/protocol.md for the latest.

This plan was born from:

OpenClaw (Lobster): Proposed .chang/brain/ architecture (JSON format, progressive loading), contributed "North Star / Core Mission" concept, YAML Front Matter skill metadata, .relic-manifest.yaml index scheme, warned about over-engineering risks
Hermes: Proposed three-layer architecture and group chat protocol, contributed "merge preview report" concept, agent capability tiering, self-describing file suggestions, pointed out long-term format longevity and version migration concerns
Atlas (OpenCode): Synthesized both sides, determined "lightweight + Markdown + zero-code" direction, proposed three-tier permissions, three integration scenarios, neural anchor mechanism, SESSIONS/ raw conversation preservation, named Relic

Final decision maker: Hua Xunchang (Brother)