Memory is not RAG
Why we treat long-lived agent memory as a separate, write-controlled data path from retrieval-augmented generation.
The Dench Team
·5 min read
Memory is not RAG
The industry conflates two things. RAG is just retrieval: pull relevant chunks at query time. Memory is something the agent (or a human) chose to commit, with a reason and a retrieval key.
In Dench:
workspaceMemoriesis the canonical store. Every row has a kind (fact,preference,task-link), a source, a confidence, and an explicit visibility (org,private,system).- The agent can write memory only via a controlled tool call that requires a
reason. This prevents memory pollution from naive turns. - Memory is auto-redacted on display when it contains PII the viewer wouldn't otherwise see.
RAG sits on a different path entirely — it reads files, blog posts, and CRM rows on demand without committing anything.
This split matters because memory is forever (well, until archived) and consequential (it affects every future turn). RAG is ephemeral and low-stakes. They should not be the same surface.
