MongoDB pushes agentic memory for enterprise AI adoption
Companies may waste spending on larger models if their data systems cannot deliver reliable context for agents
Enterprise artificial intelligence (AI) systems are running into a memory problem. Companies can buy newer models, expand context windows and launch more prototypes, but agents still falter if they cannot retain and retrieve the right business context.
That challenge is becoming more urgent as AI moves from chatbots into agents that can plan, act and query applications repeatedly. For these systems, the real bottleneck may not be the large language model (LLM) itself, but the memory and retrieval layer beneath it.
“If your retrieval isn’t first class, garbage in, garbage out. Bad AI is often less an LLM problem than it is a retrieval problem,” said Matt Johnson, field CTO at MongoDB. “The LLM is only acting on the information it’s given.”
He said companies can still get poor results with newer, more expensive LLMs if the underlying retrieval system is weak.
“If you don’t have high-quality retrieval, you’re giving it way more information on a given question than the context requires,” he said. “It’s going to spin on all that data that maybe isn’t relevant to the question.”
Johnson said agents moving beyond simple chatbots and retrieval-augmented generation (RAG) need to perceive, plan and act in a loop. Between those actions, short- and long-term memory determine whether the system can maintain trust and accuracy.
The issue is not only whether an AI system can search a database. It is whether it can understand intent, narrow the context and avoid feeding irrelevant material into a model’s answer.
He said retrieval systems need enough context to distinguish between similar or ambiguous requests. Without that extra context, an agent may return a broader set of results than the business question requires.
That problem reflects a wider shift in enterprise AI. As companies move from POC projects to production systems, agentic memory becomes a commercial issue because poor context can waste model spend, erode user trust and slow deployment.
MongoDB is a New York-headquartered database software company that helps enterprises build and run applications across cloud, on-premises and hybrid environments. The company has been expanding the platform for AI workloads through vector search, embedding models and agent-memory capabilities.
Memory costs bite
The discussion took place during a MongoDB media briefing in London on May 6. Johnson and Genevieve Broadhead, global industry lead at MongoDB, told TechJournal.uk that enterprises are entering a new phase in which agents need stronger data infrastructure to work in production.
“Without good persistent memory, agents just don’t live up to the hype,” Johnson said.
He said that agentic memory can also serve as a cost-control tool. Many LLM providers charge more for larger context windows. Putting only relevant information into that window can reduce the price per request and stretch the same budget across more interactions.
“You pay almost double with most LLMs if you need a larger context window,” he said. “If we can get the right data into that context window using a more intelligent agentic memory, you are effectively reducing the price per request.”
Developers are already using graph queries and vector lookups to prevent coding agents from rereading large parts of a codebase before every prompt or plan.
On May 7, MongoDB announced a package of updates to help enterprises run AI agents in production. The announcement included several updates:
Automated Voyage AI Embeddings in MongoDB Atlas Vector Search, now in public preview. The tool automatically creates embeddings when data is written or updated. Embeddings are numerical representations of meaning that help agents find relevant context.
LangGraph.js Long-Term Memory Store is now generally available for JavaScript and TypeScript developers. It gives agents persistent cross-conversation memory backed by MongoDB Atlas.
MongoDB 8.3, which the company said delivers up to 45% more reads, 35% more writes, 15% more ACID (atomicity, consistency, isolation and durability) transactions and 30% more complex operations over MongoDB 8.0 without changing application code.
Expanded deployment options across Amazon Web Services (AWS), Google Cloud, Microsoft Azure, on-premises and hybrid environments. Cross-region connectivity for AWS PrivateLink is now generally available as well.
CJ Desai, president and chief executive officer of MongoDB, said production agents depend on the data layer beneath them.
“The hardest part of running agents in production isn’t the model. It’s the data layer underneath it,” he said. “To trust an agent at scale, it has to retrieve the right context, hold memory across sessions and operate at machine speed, wherever the enterprise needs it.”
The pressure on databases is increasing as usage patterns change. Databases were once systems of record, then systems of engagement for customer- and staff-facing applications. Agentic AI is pushing them toward systems of action.
“We’re entering this agentic world of systems of action, where agentic workloads are completely different from what they were before,” Broadhead said. “The way customers are querying is changing. When I speak to e-commerce or travel companies, people have stopped putting in two words. They’ve started putting in seven words.”
That shift makes capacity planning harder. Human traffic usually has patterns, but agents can query repeatedly, run continuously and trigger new search and transaction loads.
“Agentic workloads are completely unpredictable,” Broadhead said, adding that retailers are only beginning to adopt agentic commerce and automated transactions, making database performance harder to forecast.
Cloud lock-in risk
Cloud lock-in is another pressure point. AI teams may build prototypes around a single model or cloud provider, only to find that the model they later need is elsewhere. That can leave customer data, order memory and agent context fragmented across systems.
“If you need to use a Google Cloud flagship AI product and your data is locked into AWS, that is going to be no good,” Johnson said.
Deployment flexibility is being driven by compliance, European data sovereignty concerns, latency, and the need to keep data close to applications. MongoDB can run on a developer laptop, in a company data center, in a virtual private cloud or through Atlas as a managed service.
Asked how MongoDB stands out from competitors in agentic memory, Broadhead pointed to its deployment flexibility rather than to a single model or a cloud tie-up.
“The place where MongoDB stands apart from competitors is the run-anywhere thing,” Broadhead said. “A lot of AI and vector databases are only available in one cloud or with one cloud provider.”
She said companies can struggle to move AI workloads into production when they build with one model and then discover that the model they need runs on another cloud provider.
Security is also becoming more complicated. Johnson said agentic AI makes business logic less deterministic, meaning some guardrails may have to move closer to the database or into the database itself. Sensitive fields, such as personally identifiable information (PII), may require controls to prevent agents from accessing or releasing data they should not see.
“Customers don’t really buy products. They buy trust,” Johnson said.
He said companies were not discussing agentic memory and production AI architecture 18 to 21 months ago. Customers now ask about agents, models, clouds and guardrails before they get to the database discussion.
“Anyone who says they’re an AI expert is completely lying to you,” Broadhead said. “Maybe they were yesterday, but not today.”
She said MongoDB itself has had to widen its role as customers ask more questions about AI architecture, agents, models and guardrails.
For enterprises, the next phase of production AI will depend less on isolated experiments and more on disciplined data architecture. Agents will need reliable memory, flexible deployment and security controls before they can be trusted to act at scale.