Memory vs RAG
They're different jobs. RAG finds passages in your documents; memory maintains what's currently true about a user or project. RememberOS does both, over the same pool.
| RAG | Memory | |
|---|---|---|
| Unit | document chunks | atomic facts |
| Writes | add, file drops, connectors | remember (LLM extraction) |
| Time | static — a chunk never changes | evolving — facts supersede, extend, expire |
| Read | search (hybrid) | search (current-truth) + profile |
| Cost | no LLM calls | LLM calls per remember |
Use RAG when…#
- the source of truth is documents (manuals, transcripts, knowledge bases);
- you want passages with provenance, not conclusions;
- volume is high and per-item LLM cost would hurt.
Use memory when…#
- you're tracking users, customers, or projects whose state changes;
- contradictions must resolve ("works at Google" → "works at Stripe");
- you want a profile you can drop into a system prompt.
Use both (the usual answer)#
Drop the user's documents (RAG) and remember what conversations
reveal (memory) into the same collection — one search spans both, and the
profile captures the durable facts. The pool is shared; the indexing differs.