Install Everything

Full System Walkthrough

Open each section as you go. The progress meter updates as the stack unfolds, from raw memory storage to the HARVEST loop that turns yesterday's sessions into tomorrow's context.

Prerequisites

Start with an agent host

OpenClaw is the platform layer shown here most prominently, but Hermes Agent is another important agent host in the same family of workflows.

Get OpenClaw → Get Hermes Agent →
GitHub: openclaw/openclaw GitHub: NousResearch/hermes-agent

Resources

Everything you need before you scroll

Open the exact package, reference, or platform link for each layer before you work through the full walkthrough.

1. MemPalace

Get MemPalace →

Install it first, then point it at your notes, transcripts, and project folders.

GitHub npm

2. LLM Wiki Pattern

Read the Pattern →

Use the original write-up as the concept anchor, then express the structure inside your own WIKI.md.

Original LLM Wiki gist

3. EJ's Palace Wiki

Browse EJ's GitHub Org →

Reference EJ's public org while you build the wiki-centered variation of the pattern.

Open FuFicFac

4. SOUL.md + MEMORY.md

Create the Context Files →

Create these files yourself, using existing agent context repos as the reference pattern. OpenClaw and Hermes are both worth studying if you want persistent agents rather than one-off chats.

Claude Code context OpenClaw docs Hermes Agent

5. HARVEST

Read the HARVEST Pattern →

HARVEST is documented in the EJ Palace Wiki pattern. Use that reference first, then run the scheduled learning loop on OpenClaw, Hermes, cron, or whatever host fits your system.

HARVEST pattern Companion repo

1. MemPalace

Install durable semantic memory before anything else.

Expand

MemPalace is the raw memory substrate. It stores source material, session logs, notes, and transcripts as searchable vector-indexed drawers so an agent can retrieve prior context instead of rebuilding the world every time you open a terminal. 

npm install -g mempalace

# create mempalace.yaml and point it at your source directories
mempalace ingest ./Inbox
mempalace query "What did we learn from yesterday's session?"

The headline reason this matters is continuity. EJ's production memory has 43,000+ drawers, which means the agent starts from accumulated reality, not a blank whiteboard.

2. Karpathy's LLM Wiki Pattern

Curate the important 5% into pages the model can maintain.

Expand

The wiki pattern keeps the raw archive from becoming a swamp. Most information stays raw; the best 5% gets promoted into a living wiki through four operations: INGEST for new material, QUERY for retrieval, HARVEST for missed facts, and LINT for structural health checks. 

INGEST  -> process notes and sources into candidate knowledge
QUERY   -> search the wiki and answer from curated context
HARVEST -> pull durable facts out of real work sessions
LINT    -> find broken links, stale pages, and schema drift

That is what makes the system alive. A static wiki decays. An operated wiki keeps being re-shaped by use.

3. EJ's Palace Wiki

Bridge MemPalace and the curated wiki into one operating system.

Expand

EJ's variation adds HARVEST directly into the wiki workflow. MemPalace catches what live sessions produced, HARVEST turns the durable parts into candidates, and the wiki absorbs the results so future agents inherit them by default. 

WIKI.md      -> schema, voice, and editorial rules
index.md     -> catalog of important pages
log.md       -> operational log of what changed
Wiki/        -> Projects, People, Concepts, Infrastructure
Inbox/       -> drop zone for unsorted material

Without HARVEST, sessions are islands. With it, sessions become source material for the next round of thinking.

4. SOUL.md + MEMORY.md

Give any model a stable identity and a working sense of the room.

Expand

These files are the agent's onboarding packet. SOUL.md defines personality, values, voice, and guardrails. MEMORY.md captures standing facts, active context, and operational rules. Any capable model can read them, whether that is GLM 5.1, GPT-5.4, or Claude Opus. 

SOUL.md
- who the agent is
- how it should sound
- what it protects

MEMORY.md
- standing context
- ongoing projects
- facts worth keeping between sessions

This is the move that makes the agent model-agnostic instead of provider-locked.

5. HARVEST Deep Dive

Automate the learning loop while you sleep.

Expand

HARVEST runs on a schedule, mines agent sessions, scores candidate facts, then routes them by confidence: above 0.85 gets auto-promoted, 0.5 to 0.85 becomes a review candidate, and anything lower gets discarded. 

Work
  -> Harvest
    -> Wiki
      -> Context
        -> Better Work
          -> Harvest again

That is the difference between a file cabinet and a living system. The loop keeps folding real work back into usable context.

You're Done

The stack is ready to compound

Next step: ingest reality

Point MemPalace and the wiki inbox at your real notes, transcripts, bookmarks, and active project folders.

Next step: schedule HARVEST

Run the first loop on recent sessions so the system demonstrates improvement immediately instead of waiting weeks.

Walk it step by step Back to guide home