EJ Palace Wiki / MemPalace

Personal Knowledge Architecture

Wiki = curated 5%. MemPalace = searchable 100%.

You

🧠

You

Questions · Files · Sessions · Decisions · Notes

drop a file ask a question end a session make a decision

Inbox/

DROP ZONE

New sources arrive here before processing.
LLM picks them up and decides: does this update a durable entity, create a wiki page, or go straight to Raw Sources?

Operations

Ingest

→New source arrives →LLM reads it →Updates durable entity? →Wiki pages updated →Source → Raw Sources →MemPalace indexes it

Query

→User asks a question →LLM reads index.md →Reads relevant pages →Searches MemPalace →Synthesizes answer →Files evergreen facts back

Harvest

→Scheduled job runs →Mines agent sessions →Extracts wiki-worthy facts →Promotes into shared wiki →Bridges isolated agent knowledge → shared wiki

Lint

→Periodic health check →Orphan pages →Broken links →Stale pages →Contradictions ⚑No auto-fix w/o approval

Core Two-Layer System

Wiki

Curated Knowledge

~5%

LLM-maintained markdown pages. Cross-linked, human-readable synthesis. Durable entities, active projects, real people, recurring concepts, infrastructure, and decisions with reasoning.

Target steady state: 200–400 active pages

WIKI.md — schema + operating manual
index.md — catalog of all pages
log.md — append-only operation log

Wiki/
  Projects/
  People/
  Concepts/
  Infrastructure/
  Archive/

MemPalace

Semantic Substrate

100%

Everything, indexed. Semantic search across raw sources, session logs, notes, transcripts, and long-tail memory. Accessed through CLI or MCP. The recall layer that Wiki synthesis draws from.

Raw Sources Session Logs Notes Transcripts Decisions Long-tail memory Semantic Vectors CLI + MCP access

          Raw Sources/

            YYYY-MM/  ←  archive of ingested files

Any LLM with File Access

Consumers — read Wiki first, search MemPalace for long-tail recall, maintain both over time

🖥 Claude Code

⚡ Codex

🤖 Claude Desktop

🐾 OpenClaw

🪄 Cursor

🌐 ChatGPT / MCP clients