Governed memory vs RAG

RAG (retrieval-augmented generation) is a technique: fetch relevant chunks, stuff them into a prompt, generate an answer. Governed memory is an operating model: the organization decides what may be remembered, who may read or write it, how freshness is proven, and how agents may use it.

You almost always need retrieval. You do not automatically get a company brain by embedding your drive.

What RAG optimizes

RAG systems optimize for relevance under latency and cost constraints:

  • Chunking and embedding corpora
  • Vector / hybrid search
  • Context packing and citation of source documents
  • Sometimes reranking and query rewriting

Done well, RAG reduces hallucinations relative to a naked LLM. Done alone, it still answers from whatever text was indexed — including outdated, conflicting, or over-permissioned material.

What governed memory adds

Governed memory assumes retrieval exists and asks harder questions:

  1. Is this claim still true? Freshness SLAs, review dates, and retirement — not only “indexed last night.”
  2. Who owns it? Accountability for correctness when an agent relies on it.
  3. Who may see it? Access control aligned with systems of record, including agent identities.
  4. How did it get here? Provenance from source systems, humans, or promotion workflows.
  5. May the agent write back? Controlled promotion of new facts, not silent vector pollution.
  6. Can we audit use? Logs of what memory influenced which action.

That stack is what turns “search over docs” into company brain behavior.

Common failure modes of RAG-only setups

  • Stale high-rank chunks win because they are well-written, not because they are current.
  • Secret bleed via over-broad index permissions or copied text in shared docs.
  • Conflicting truths with no resolution policy — the model averages or picks arbitrarily.
  • Write amplification — agents or users dump chat summaries into the index with no review.
  • False confidence — fluent answers with citations to the wrong era of policy.

How they fit together

A healthy architecture often looks like:

  • Systems of record (code, CRM, billing) remain authoritative.
  • Governed memory holds durable explanations, decisions, playbooks, and links to those systems.
  • RAG / search retrieves across governed records and long-tail docs, with governed records preferred for high-stakes paths.
  • Agents call tools that enforce policy on read/write, not raw “dump top-k into prompt.”

Mensara and similar products emphasize the governed layer — so RAG is a capability inside a trust model, not the whole product.

Operator checklist

When someone says “we have RAG,” ask:

  • Can we list owners and last-verified dates for the top 50 agent-used claims?
  • Can we revoke an agent’s access to a memory class without reindexing everything by hand?
  • Is there a promotion path from Slack/ticket → canonical record?
  • Do evaluations measure correctness under drift, not only retrieval recall?

If the answers are thin, you have retrieval — not governed memory. Continue with Agent memory access control and Company brain freshness and drift patterns in this directory’s build guides.