Semantic recall is the embedding-based retrieval step that finds memories relevant to the current message, even when the wording has changed since the original fact was stored.
Last verified: April 2026
What is Semantic Recall?
Semantic recall is the embedding-based retrieval step that finds memories relevant to the current message, even when the wording has changed since the original fact was stored.
On each inbound message the soul loader embeds the user text with OpenAI's text-embedding-3-small, then runs a pgvector cosine distance search against user_memory, returning the top ten non-structural facts. Trivial messages (short greetings, single emoji, stickers) skip the recall entirely to save tokens and latency.
FAQ
Semantic Recall FAQ.
Common questions about Semantic Recall in the AGNT platform.
Semantic recall is the embedding-based retrieval step that finds memories relevant to the current message, even when the wording has changed since the original fact was stored.
People also ask.
Related terms
User Memory
User memory is AGNT's per-user fact store: key-value pairs with confidence scores and pgvector embeddings, split into structural keys that always enter the prompt and semantic keys recalled on demand.
Soul Loader
The Soul Loader is AGNT's system-prompt builder: it assembles the active prompt from user memory, venue soul, time-of-day context, and recent activity on every turn.
pgvector
pgvector is the PostgreSQL extension that adds a vector data type and approximate-nearest-neighbour search, letting AGNT store embeddings next to relational rows in one ACID transaction.
Structural Keys
Structural keys are the user-memory fields that are always included in the system prompt — diet, interests, last_booking, typical_party_size, fitness_goal, and a dozen more — so they never require embedding search.
Source: agnt-backend/app/core/soul_loader.py
Share as social post
What is Semantic Recall? Semantic recall is the embedding-based retrieval step that finds memories relevant to the current message, even when the wording has changed since the original fact was stored. https://agntdot.com/glossary/semantic-recall
247 / 280 chars
See it in action.
Now you know what Semantic Recall means. Try the live scan demo or read the developer docs to go deeper.