Semantic Recall: Helping Your AI Find What It Saved

You saved the definition last week. Today, your AI asks you to explain it again.

The definition may still be there. Your agent just failed to find it.

That is the problem behind our latest ClariLayer update. Semantic recall is now generally available in the hosted service, with explicit organization consent. It helps related wording surface relevant saved context, alongside the keyword matching that already exists. Automatic indexing builds and refreshes the search index as you add or edit eligible content.

Why remembering was only part of the job

Consider a saved definition called “Customer churn.” It explains that an account counts as churned only after its paid subscription ends. A cancellation scheduled for next month does not count yet.

Later, you ask your agent: “Which customers have we lost this quarter?”

This is an illustrative example, but the retrieval problem is familiar. The question and the saved entry use different words. A search that relies heavily on matching names can miss a relevant definition even though the information is already stored.

The agent may ask you to explain the rule again. Or it may proceed without the distinction between a scheduled cancellation and an ended subscription. You are back to correcting something you already taught it.

We built this update to reduce those misses. Saving context should make it easier to use in the next conversation, even when you phrase the question differently.

Recall can now match meaning as well as wording

Semantic recall adds another way to find related context. It uses embeddings: numerical representations of eligible saved text and the recall query, which help identify related meaning even when the words differ.

Keyword matching stays in place, and exact-name matches remain protected. The semantic lane contributes a bounded set of additional candidates. Existing access permissions still apply.

In the churn example, this gives the saved definition another route into the agent's recall results. The agent still needs to read the definition and apply its actual rule. Related wording does not make two business definitions interchangeable.

Recall continues to return context previews. When the agent needs a complete definition or saved SQL, it fetches the full entry separately.

The index updates as your context changes

After an owner or admin grants consent, eligible existing context is queued for indexing automatically. New and edited entries are queued for background updates too. You do not need to run a manual indexing job each time you save something.

Updates are asynchronous. Granting consent does not instantly make every entry searchable by meaning. Settings shows indexing progress, skipped entries and failures that can be retried. Entries outside the indexing limits remain available to keyword recall, and provider failures fall back to keyword recall.

This keeps the maintenance work out of the normal conversation. You save a corrected definition; ClariLayer queues the corresponding search update.

More useful memory for the next question

The practical improvement is a better chance of bringing the context you already saved into the work you are doing now.

You can ask in the language of the task instead of trying to remember the exact title of a note. Your agent has another way to find a relevant business rule before it starts answering. As your saved context grows, that ability becomes more useful.

Retrieval is the part we improved here. A relevant match does not establish that a definition is correct or current, and this release does not guarantee a correct AI answer. The saved entry's provenance and trust status still matter.

Available now, with an explicit choice

Any organization owner or admin can enable the feature without a manual allowlist request.

Open Settings → Account, find Semantic indexing, and choose Grant consent. Indexing then runs in the background.

An owner or admin entering Connect your AI for an organization that has not decided will also see Enable semantic search and Not now. Not now records the organization's choice to leave it off; an owner or admin can enable it later in Settings. Existing decisions are preserved. Signing up or accepting the Terms of Service does not grant consent.

Enabling this allows eligible saved entry names and main text, including existing and future context, plus recall query text, to be sent through Vercel AI Gateway to Google to create embeddings. ClariLayer stores the resulting entry vectors. This choice applies to the organization.

You can revoke consent in Settings. Revoking removes the organization's stored embedding rows while keeping the saved context. Requests already in progress may finish, and data already sent cannot be recalled. The semantic indexing privacy notice explains the data flow.

Key takeaways

• Semantic recall helps related wording find relevant saved context while retaining keyword matching.
• Indexing and updates run automatically after explicit consent; new accounts are not opted in by default.

Enable semantic indexing in Settings, then keep working in the AI you already use.