AI · Memory
Graph Memory on VegaDūta
Graph memory is how a VegaDūta agent remembers you next week. Instead of stuffing past transcripts into the prompt, the platform extracts entities and facts from conversations into a knowledge graph — subject-relation-object triples — and agents recall from that graph as a tool when a conversation needs it.
Write arbitration keeps the graph clean as new facts arrive, and an optional tenant-shared graph lets all of a team's agents draw on the same pool of facts.
Conversations become facts
As agents talk, an extraction step identifies entities and the relationships between them and writes them into the graph as triples. A customer's preferences, an order's status, a project's owner — these become structured facts rather than sentences buried in a transcript.
Recall is a tool, not a prompt dump
When an agent needs context, it queries the graph through a recall tool and gets back the relevant facts. This keeps prompts small and answers grounded: the agent retrieves the three facts that matter instead of rereading every previous conversation.
Write arbitration keeps memory clean
Long-term memory rots if every utterance is stored verbatim. Write arbitration decides what actually enters the graph — reconciling new information against what is already there — so the graph stays a curated set of facts rather than an append-only log of everything ever said.
Shared memory for teams
Optionally, a tenant can enable a shared graph so facts learned by one agent are recallable by the others. Your sales agent and your support agent then agree on who the customer is and what was promised, instead of each keeping a private, diverging picture.
Frequently asked questions
How does an AI agent remember past conversations?
On VegaDūta, facts are extracted from conversations into a knowledge graph as entity-relation triples. In later conversations the agent queries that graph through a recall tool, retrieving just the relevant facts instead of replaying old transcripts.
What is graph memory, compared to a vector database?
A vector database retrieves chunks of text that are semantically similar to a query; a knowledge graph stores discrete facts and their relationships. VegaDūta uses both — RAG over your documents for reference material, graph memory for structured facts about people, things, and commitments learned in conversation.
Can multiple agents share the same memory?
Yes, optionally. A tenant-shared graph lets all agents on a team read the same facts, so knowledge learned by one agent is available to the rest. It is a per-tenant setting, not the default.
What stops the memory filling up with junk?
Write arbitration. New candidate facts are reconciled against the existing graph before being written, so duplicates, contradictions, and trivia are filtered rather than accumulated. The result is a curated fact store, not an append-only transcript log.
See it working in two minutes
The sandbox provisions a real tenant — describe an agent in one sentence and test it, no account, no card. Or browse ~90 industry workflow recipes to see what teams build.
