Where memory
approaches differ.
RAG, vector stores, session state, and persistent memory solve different problems. This is an educational comparison of what each approach does and where it stops.
Five approaches,
five scopes.
| Approach | What it does | Where it stops |
|---|---|---|
| Stateless prompting | Provides context only for the current request. | Nothing persists. Every session starts from zero. |
| Retrieval-augmented generation (RAG) | Retrieves selected information and places it into the model's context for a response. | Supplies context for a request; it does not by itself retain knowledge across sessions. |
| Vector store / embeddings only | Stores content as embeddings and retrieves by similarity. | Similarity is not governance. There is no claim lifecycle, provenance, or conflict handling by default. |
| Session-only memory | Holds working context for the duration of a session. | The context expires with the session. Durable knowledge is lost. |
| Governed persistent memory | Retains knowledge across sessions with scoped recall, provenance, and a claim lifecycle. | Requires defining memory policies and the deployment boundaries you intend. |
Context for a request.
Knowledge across sessions.
Retrieval-augmented generation is a pattern for supplying context to a single response. It retrieves relevant information and places it into the model's context window. That is genuinely useful, and it is not the same thing as memory.
Persistent memory retains knowledge so it remains available after the request ends and after the session closes. A decision made in one session can be recalled in the next. A constraint recorded once does not need to be re-derived.
The two are complementary. Retrieval can bring stored knowledge into a request, while persistent memory is what makes that knowledge exist in the first place. Recall Racoon uses hybrid retrieval over a governed, persistent store rather than treating retrieval as the whole answer.
Storage is not
the hard part.
Storing content is straightforward. Knowing whether it is still true is not. A memory system without governance treats every stored item as equally current, which is how stale guidance and conflicting claims end up influencing an agent.
Governance adds a claim lifecycle, so knowledge can be marked candidate, confirmed, stale, superseded, conflicted, or archived. It adds provenance, so a claim can be traced to its source. It adds conflict review, so disagreements are surfaced rather than silently resolved. And it adds scoped recall, so an agent retrieves from the intended project rather than a shared pool.
That combination is what separates a governed memory layer from a store of embeddings. The store remembers. The governance decides what deserves to be trusted.
See governed memory
in practice.
Explore the product capabilities, or read the glossary for the concepts behind them.