Agents with memory: how Agno and SurrealDB enable reliable AI systems

Source: Dev.to

SurrealDB CEO Tobie Morgan Hitchcock recently hosted Agno founder and CEO Ashpreet Singh, together with SurrealDB Solutions Engineer Martin Schaer, for a livestream to discuss how to build agents that remember, reliably. The conversation covered the realities of building production‑grade agent systems, why context is more important than raw model power, and how Agno’s Agent OS and SurrealDB fit together as a modern agent‑memory stack. This article distills that livestream into a focused Q&A for developers and teams who want their agents to do more than just respond to a single prompt.

What is Agno, and why does it pair so well with SurrealDB?

Tobie: For anyone who has not used Agno yet, what does it do, and how does SurrealDB fit in?

Ashpreet:

• Agno is a high‑performance multi‑agent framework. If you want to build serious agent systems, Agno gives you three things at once:

  1. A harness where your language models and agents run.
  2. A runtime that you can deploy in your own cloud so you keep control of your data.
  3. A user interface that lets you observe and manage your agents.

• When you build systems with LLMs you always have two pieces:

  1. The harness (Agno’s focus) – the environment where the model and agents execute.
  2. The context you feed into that harness – that’s where SurrealDB comes in.

• Agno provides the Agent OS for orchestrating agents.

• SurrealDB provides the memory and knowledge layer that feeds those agents with the right information.

• From the beginning, the integration between Agno and SurrealDB was driven by users who were already combining both. They needed a way to run multi‑agent workflows and a database that could handle memory, knowledge, and graph relations at the same time. That combination is what this partnership is about.

What are teams struggling with when they build agents?

Tobie: As developers and companies build on Agno, what are the biggest challenges you are seeing?

Ashpreet: We see two big challenges again and again.

1. Reliability and visibility – Teams want to know whether what they have built actually works. They ask:

   • How is the system performing?
   • When can this go live?
   • What happened in this specific interaction?

   It is not enough to have something that works sometimes. They need repeatable behaviour and a way to inspect it.

2. Context – With language models, the main thing you control is the input. If the context is wrong or noisy, the output will be unpredictable. Teams struggle with how to consistently provide the most relevant and up‑to‑date context, so users get a good experience every time instead of occasionally.

Is the answer just bigger context windows?

Tobie: Many developers try to fix this by throwing more context at the model. Does that work?

Ashpreet: Not really. People are excited about huge context windows, but in practice you do not want to send everything you have on every request. If you have a million tokens available, you should not send a million tokens. You should send exactly what you need.

Two points matter here:

• It helps to use the best model you can for the task, but even then, sending too much context makes the model less focused and increases cost.
• Consistency does not come from size. It comes from giving the model tightly scoped information that is clearly relevant to the task.

This is where a proper memory and knowledge architecture is important. You want to store a lot, but you want to retrieve precisely. That is the role of SurrealDB in this stack: a single place that holds structured facts, vectors, relationships, and history, which Agno can query in a controlled way.

How exactly does SurrealDB integrate with Agno’s Agent OS?

Tobie: What did the Agno × SurrealDB integration actually add for developers?

Martin: The integration arrived in two phases.

1. Vector‑store integration – We first integrated SurrealDB as a vector store for knowledge. This allowed Agno agents to perform semantic search across documents stored in SurrealDB and use that as their knowledge base.

2. Native memory provider – More recently we merged SurrealDB as a native memory provider inside Agent OS. Now agents can store sessions, memories, knowledge snapshots, evals, and metrics directly in SurrealDB. For each agent run you can see, in the database, how many tokens were used, how long it took, what context was fetched, and what was stored as memory.

In practice: Agno agents use SurrealDB both as long‑term memory and as a source of truth for analytics around their own behaviour. This makes debugging and improving agents much easier, because you can inspect what actually happened rather than guessing.

How does Agno think about “context” and “memory”?

Tobie: The industry uses many terms like context, knowledge, memory, brain, and reasoning. How do you think about these inside Agno?

Ashpreet: We try to avoid buzzwords and think of everything as context. By context we mean the input that is available to the program that is calling the model. You can break that down into three parts.

1. Task prompt – The description of what the agent should do.

   • Example: “You are answering user‑support tickets and must respond in a friendly, concise style.”

2. Memory – Information about specific users, sessions, or previous interactions.

   • Example: “User lives in New York, likes product X, prefers email communication.”
   • This is where SurrealDB is often used as a memory store.

3. Knowledge – Your documents, past tickets, product documentation, research notes, etc.

   • SurrealDB acts as the backing store, using vectors, documents, and graphs.

The important design choice is that the model is not simply force‑fed everything all the time. Instead, the agent is given tools it can use to retrieve the right context on demand.

Why is SurrealDB a good fit for agent memory and context?

Tobie: Many storage systems can hold vectors. What makes SurrealDB useful as a memory layer for agents?

Martin: In practice, agents rarely need only embeddings. Real workloads involve multiple types of data at once.

• Teams use SurrealDB for structured facts, vectors, graph relationships, and time‑series data.
• Typical pattern: store users and accounts, connect them via graphs, attach embeddings to content, and record timestamped events.

Retrieval strategy:

1. Similarity search over vector data.
2. Refine results using graph relations (e.g., filter by tenant, time window, or follow relationships to accounts, tickets, documents).
3. Time‑awareness – each chunk can be associated with a month or period, allowing you to keep only the relevant timeframe.

The result is a cleaner, more relevant context for the agent. Instead of “top N vector matches”, the agent sees a carefully curated slice of the user’s actual world, improving accuracy and making the system more deterministic.

What are people actually building with Agno and agent memory today?

Tobie: In production environments, what kinds of agent use cases are working well right now?

Ashpreet: The majority of successful use cases look quite practical—and sometimes even boring—but they deliver real value.

Common patterns

1. Document processing & extraction pipelines

   • Invoice processing, contract analysis, key‑field extraction from semi‑structured or visual documents.
   • Multimodal: PDFs, images, and text together.

2. Internal knowledge assistants

   • Help prepare for sales calls, summarise incidents, surface relevant documents, and pull together history around specific customers or features.

3. Support workflows

   • Agents read historical tickets, understand user history, examine screenshots or charts, then draft responses or suggest next actions.
   • Multimodal models + strong memory layer shine here.

In almost all of these cases, SurrealDB is the central store of knowledge and memory, while Agno provides the harness and tools to orchestrate the agents that sit on top of that data.

Are we close to fully autonomous, company‑wide agents?

Tobie: Many organisations dream of one large agent that can roam across all company data. How realistic is that today?

Ashpreet: We are not there yet, and it is important to be honest about that.

• Current state: Impressive demos and highly effective agents for specific workflows are possible.
• Full autonomy across all data & workflows is still beyond the current ecosystem.

How the situation improves

1. Better models – continual advances in LLM capabilities.
2. Maturing harnesses & systems – e.g., Agent OS, richer memory/data layers like SurrealDB, and stronger engineering practices.

Practical advice:

• Build small, tightly scoped agents that solve clearly defined problems extremely well.
• Once you have a solid harness and a strong memory layer, you can grow those agents into larger workflows as the technology matures.

How should developers get started with Agno and SurrealDB?

Tobie: For a developer who has only used chat‑style interactions so far, what is a good first step into agents with memory?

Ashpreet: Start small and start with a problem that matters to you.

1. Pick a personal or internal workflow you understand deeply (e.g., a notes assistant, daily message summary, or internal helper for company docs).
2. Use the Agno documentation and the SurrealDB integration cookbook to spin up Agent OS connected to SurrealDB.
   • Let SurrealDB store your knowledge, sessions, and metrics.
   • Let Agno orchestrate the agents and tools.

Key tip: Don’t overthink the first version. The goal is to feel how the harness, the memory, and the model interact. As you iterate, you’ll discover that improvements often come more from better retrieval and memory design than from simply switching to a larger model.

What is coming next for agent memory with Agno and SurrealDB?

Tobie: Looking ahead, what is most exciting about the future of Agno and SurrealDB in the agent space?

Ashpreet: One idea we are very excited about is the notion of Agent OS as a networked runtime.

• Agent OS can expose agents in multiple ways, including as MCP tools and as APIs.
• Inside a company you can imagine many different agent runtimes, each owned by a team and specialised on a certain domain.
• These agents can talk to each other, with SurrealDB acting as the shared or scoped memory layer underneath them.

SurrealDB roadmap

• Richer multimodal support – handling images, audio, video alongside text.
• Unified abstractions for memory across facts, vectors, graphs, and time‑series.
• Goal: Agents simply describe the kind of information they need, and the combination of Agno and SurrealDB will handle fetching and shaping that context.

Bottom line: Reliable agents need a strong harness and a strong memory layer. Agno focuses on the harness; SurrealDB provides the memory. Together they are paving the way toward more capable, interconnected agents.

Harness

SurrealDB focuses on memory and context. Together they help teams move from a “cool demo” to real, production‑grade agents that actually remember.

Start building

• Create your SurrealDB account – get a ready‑to‑use free environment for storing agent memory, knowledge, and context.
  👉

• Check out the Agno and SurrealDB integration – see how Agent OS uses SurrealDB for memory, sessions, evals, and knowledge retrieval.
  👉

• Explore the new SurrealDB integration inside Agno – use the cookbook examples to spin up agents, test memory workflows, and experiment with context‑rich behaviours.
  👉