← Back to EJ's Demo Sites
5%
Personal Knowledge Architecture · EJ Palace

Wiki
meets
MemPalace

Wiki = curated 5%. MemPalace = searchable 100%.
Two layers. One brain.

🧠
You
questions
files
sessions
decisions
notes
Inbox/
drop zone · LLM reads each file · decides: update wiki entity? create page? index only?
// Operations — the four verbs
01
INGEST
source arrives
LLM reads it
durable entity?
wiki updated
→ Raw Sources
MemPalace indexes
02
QUERY
question arrives
reads index.md
reads pages
searches MemPalace
synthesizes answer
files evergreen facts
03
HARVEST
scheduled run
mines agent sessions
extracts wiki-worthy
promotes facts
bridges isolated
→ shared wiki
04
LINT
periodic check
orphan pages
broken links
stale pages
contradictions
⚑ no auto-fix
Wiki
~5%
Curated.
Durable. Linked.
LLM-maintained markdown. Cross-linked pages. Human-readable synthesis. Durable entities, active projects, real people, recurring concepts, important decisions with reasoning.

Target: 200–400 active pages.
WIKI.md  schema + manual
index.md  full catalog
log.md    append-only log

Wiki/
  Projects/
  People/
  Concepts/
  Infrastructure/
  Archive/
⚡
MemPalace
100%
Everything.
Searchable. Recalled.
Semantic search across the full long-tail corpus. Raw sources, session logs, notes, transcripts. Accessed via CLI or MCP. The recall substrate that wiki synthesis draws from.
raw_sources session_logs notes transcripts decisions vectors long_tail
$ mempalace search "chorus architecture"
$ mempalace query --wiki-first
RAW SOURCES/
YYYY-MM/ ← ingested files archived here after INGEST runs
♻
HARVEST bridges the gap: agent session memory that lives in isolation gets mined nightly and promoted into the shared wiki — so knowledge built by Butch, Alice, Hank, and Claude agents stays synchronized.
// Any LLM with file access — reads wiki first · searches MemPalace for long-tail · maintains both
🖥
CLAUDE CODE
⚡
CODEX
🤖
CLAUDE DESKTOP
🐾
OPENCLAW
🪄
CURSOR
🌐
MCP CLIENTS
human input
inbox / raw
ingest
query
harvest
lint
wiki (~5%)
mempalace (100%)
consumers