Step 4

Install SOUL.md and MEMORY.md

These files give AI agents persistent identity and memory. Any capable model can read them and immediately understand who it is, how it should behave, and what facts it should carry forward.

Create These Files

Get SOUL.md + MEMORY.md

Create them in your project root, then load them every time the agent boots so identity and context stay stable. For agent-hosted workflows, reference both OpenClaw and Hermes Agent.

Get SOUL.md + MEMORY.md →OpenClaw docsHermes Agent
Reference: anthropics/claude-codePlatform: openclaw/openclawPlatform: NousResearch/hermes-agent

Overview

Why context files matter

SOUL.md defines personality, values, voice, and guardrails. MEMORY.md stores standing facts, active context, and operational rules. Together they make the system model-agnostic across GLM 5.1, GPT-5.4, and Claude Opus.

Deep Dive

Expand the file roles

This is the human-readable contract an agent reads before it starts making decisions.

What belongs in SOUL.md

Personality and behavior constraints, not project scratch notes.

Expand

SOUL.md should answer questions like: Who is this agent? What tone does it use? What values shape its decisions? What lines should it not cross? This keeps the agent coherent across sessions and across model swaps.

SOUL.md
- name and role
- voice and emotional range
- values
- collaboration style
- guardrails

What belongs in MEMORY.md

Stable facts, active context, and operational truths worth reloading.

Expand

MEMORY.md is the curated long-term context. It should track projects, important personal facts relevant to work, active commitments, and reliable lessons learned. It is not a full diary. It is the distilled state that helps the agent start informed.

  • standing project context
  • important people and relationships
  • current rules of operation
  • durable facts the model should not keep forgetting

Why this is model-agnostic

The file format travels better than a provider-specific feature.

Expand

Any model that can read text can read these files. That means the agent's identity and working context survive platform changes. You are not betting your organizational memory on a single chat product keeping state correctly forever.

GLM 5.1   -> reads SOUL.md + MEMORY.md
GPT-5.4   -> reads SOUL.md + MEMORY.md
Claude Opus -> reads SOUL.md + MEMORY.md

Same files. Different models. Consistent operator.

Installation

Step-by-step

Install just this

Do this now if your agent feels smart but inconsistent.

  1. Create SOUL.md with voice, role, values, and guardrails.
  2. Create MEMORY.md with durable facts and current operating context.
  3. Reload the agent with both files available at startup.
touch SOUL.md MEMORY.md

# SOUL.md = identity
# MEMORY.md = standing context
# refresh the agent so it reads both on boot

Install as part of the full system

Tie these files to the wiki and memory architecture so all four layers reinforce each other.

  1. Mirror major project names between MEMORY.md and the wiki.
  2. Use HARVEST outputs to propose MEMORY.md updates when appropriate.
  3. Keep SOUL.md stable and MEMORY.md curated, not bloated.
Agent boot
  -> SOUL.md
  -> MEMORY.md
  -> QUERY wiki
  -> QUERY MemPalace

Now the agent starts oriented before live work begins.
Previous: EJ's Palace WikiNext: HARVEST