Compared with RAG and LLM Wiki

CentricMem compiles knowledge when you file it — Markdown cards on a hosted librarian — instead of slicing raw documents at question time. That is the same ingest-time idea as Andrej Karpathy’s LLM Wiki. The data model is not a wiki: typed cards, append-only decisions, capture left in the agent you already use. It is not a vector RAG product.

  • CentricMem compiles knowledge at ingest time into Markdown cards. Traditional RAG slices raw documents at query time.
  • Search is BM25 on cards. Originals are human downloads. FTS does not index attachment fulltext.
  • Karpathy’s LLM Wiki shares ingest-time compilation. CentricMem does not implement entity pages, [[wikilinks]], or an Obsidian graph.
  • Decisions append-only and supersede. Capture stays in the agent. Do not dump every chat.
  • Not a RAG alternative and not an LLM Wiki. Manager layer over agent-native memory, plus a literature database.
  • Compare tables: https://centricmem.com/compare
Query-time chunk retrieval versus ingest-time Markdown cards.
DimensionTraditional RAGCentricMem
When knowledge is compiledQuery time: each question slices raw documents into chunks.Ingest time: the agent reads the source and writes a card (summary + key points).
What is searchedChunks of originals, usually by embedding.Cards on the librarian. FTS does not index attachment fulltext. Humans download originals.
How knowledge accumulatesIt does not. The next question starts from the raw corpus again.Filed cards stay. The next agent, on another desk or model, cites the same record.
ConflictsOverlapping chunks reach the model; it judges them live.Decisions append-only and supersede. Ambient warns when active decisions overlap.
CaptureThe corpus often is the product.Capture stays in Cursor memories or similar. CentricMem organises what you file.
RetrievalA vector store or vendor memory API.BM25 on Markdown cards. Not a third-party vector API.
Karpathy’s entity wiki versus typed cards on a hosted librarian.
DimensionLLM WikiCentricMem
JobPersonal or research compiler: entity and concept pages that compound as you read.Manager layer + literature database for agent-native work. Handoff is a card, not a meeting.
Compiled unitLiving wiki pages. One source may update 10–15 pages.Typed cards: note, decision, lesson, session. Every card needs a summary and key points.
Links[[wikilinks]] and an Obsidian graph.Library → named shelf → card. Tags and decision numbers (#0003).
How pages evolveRewrite concept pages so the synthesis stays current.Decisions append-only. Supersede; do not rewrite history.
What you ingestDrop sources; the LLM wiki-ifies. Karpathy prefers one-at-a-time with a human in the loop.Selective. Skip Micro. Do not dump every chat. Capture stays in the agent.
SchemaCLAUDE.md or AGENTS.md in a vault, co-evolved per domain.The centricmem-agent Skill. Same folder name on every agent.
Where it livesA local Obsidian vault, usually git.Hosted librarian. Agents upload via MCP. You download to open.
MaintenanceLint for orphans, dead links, missing concept pages, stale claims.Ambient conflict warnings, distill, promote. Isolation is shelves and keys.

FAQ

Is CentricMem a RAG product?
No. RAG slices raw documents at query time. CentricMem files Markdown cards at ingest time and searches those cards with BM25. Originals sit in object storage for humans to download. It does not replace a corpus pipeline you already run.
Is CentricMem an LLM Wiki?
It shares Karpathy’s ingest-time compilation idea: compile once when you file, then search the compiled store. It is not the gist’s entity wiki — no [[wikilinks]], no living concept pages. Typed cards, append-only decisions, hosted librarian.
Which should I use?
Keep RAG if you query a raw corpus. Use an LLM Wiki vault if you want an Obsidian entity graph. Use CentricMem when agents on different desks must cite the same audited cards without dumping every chat.

Home · Pricing · Compare · Docs · Legal notice · Terms · Privacy · Refunds · Sign in · Sign up · Public skill · llms.txt · llms-full.txt