- AWS launched its Context service on June 17, 2026, mapping enterprise data lakes and warehouses into a queryable knowledge graph for AI agents.
- Google updated its Gemini Enterprise Agent Platform around June 29-30, 2026, adding a dynamic Memory Bank that rewrites stored context as agents interact, plus a graph-based ADK for multi-agent orchestration.
- The practical choice between them comes down to whether a team needs structured access to pre-existing enterprise data (AWS Context) or a self-updating memory layer built from live agent activity (Gemini Memory Bank).
AWS and Google both shipped graph-based context layers for AI agents in the same week, each attacking the same blind spot in vector-based RAG: it retrieves similar text but cannot reliably connect related facts across time, systems or causal chains. AWS Context, launched June 17, 2026, maps existing enterprise data into a queryable knowledge graph. Google’s Gemini Enterprise Agent Platform, updated around June 29-30, pairs a dynamic Memory Bank with a graph-based Agent Development Kit (ADK) for multi-agent orchestration.
The Limitations of Unstructured Agent Memory
Most early agent architectures lean on Retrieval-Augmented Generation paired with vector databases. The agent retrieves semantically similar text chunks and injects them into the LLM’s context window. It works well enough for basic lookup and summarisation. It breaks down when the data has dependencies.
The core failure is temporal: a vector store cannot easily distinguish what is true now from what was true last quarter. Feed a procurement agent stale supplier terms and the downstream decisions get messy fast. Flat text retrieval also loses causal links. Consider a customer support scenario where the resolution requires knowing that Customer A bought Product B, which has a known bug C, that a workaround D was applied by Agent E last week. A vector search might surface fragments about each of those, but assembling them into a coherent, actionable picture requires explicit relationship representation, exactly what a graph provides. That gap is what AWS and Google are both targeting.
AWS Context: Bridging Data Silos with Knowledge Graphs
AWS automates the process of turning a company’s existing data lakes, warehouses and databases into a connected knowledge graph.
The multi-hop reasoning case is where this pays off most clearly. A security agent can traverse the graph to trace why vulnerability A connects to system compromise B, which depends on codebase C running inside application D, with users X, Y and Z at risk. Traversing that chain from raw vector retrieval alone is unreliable. From a structured graph, it becomes a query. AWS’s framing is that agents get smarter not just through better models but through access to curated, relational context at scale.
Integration centres on existing AWS data services, so teams already running on the AWS stack face less migration friction. Pricing details were not released at launch, with AWS’s typical usage-based billing for data processing, storage and API calls expected to apply. The real caveat: enterprises with weak data governance will need to invest in ontology definition before the automated mapping delivers clean results. How much effort that requires will vary considerably depending on the state of a company’s existing data infrastructure.
Gemini Enterprise’s Memory Bank: Dynamic, Evolving Agent Context
Where Google’s approach differs is in the continuous update loop. Memory Bank does not just hold a snapshot of enterprise knowledge, it rewrites stored memories as agents interact, persisting user preferences and session history across conversations. An agent working with a returning customer can draw on what it learned three sessions ago, not just what it retrieved from a static corpus.
The updated ADK adds a graph-based framework for orchestrating networks of sub-agents. That matters for workflows like e-commerce order resolution, where a master agent needs to coordinate inventory, customer service and finance agents without losing thread across the handoffs. Memory Bank also handles multimodal inputs and supports asynchronous memory generation, so the agent keeps working rather than pausing for the memory write to complete.
On the operational side, Gemini Enterprise’s Agent Runtime supports long-running tasks with efficient provisioning and fast cold starts. Security tooling includes Agent Identity for traceability, Agent Gateway for controlled interactions and Model Armor for prompt injection defence.
Why Graph-Based Memory Matters for Enterprise AI
The reasoning and explainability gains are the most immediate argument. When a graph models entities and relationships as nodes and edges, causal chains and temporal context survive in a queryable form. A financial risk agent can explain which facts and policies led to a given recommendation, tracing the lineage step by step. That audit trail is what compliance and governance teams actually need: not just an answer, but the reasoning behind it.
Persistence is the second factor. Standard context windows are finite; agents forget. Graph-based frameworks like Zep‘s Graphiti, which attaches temporal validity windows to individual facts, let agents distinguish current from historical data without loading entire conversation histories into the prompt on every call. Long-term personalisation, remembering a user’s preferences across weeks of interactions, becomes tractable.
Cross-system integration is where the architecture earns its place in enterprise stacks. Enterprise data lives in CRMs, ERPs, data lakes and operational databases that rarely talk to each other cleanly. A context graph sits across those silos and gives agents a unified query surface. A sales agent could pull customer history from Salesforce, order detail from an ERP and recent communication context, then reason over the combined picture rather than each source in isolation. That kind of cross-system agent capability is increasingly what enterprise deployments actually require.
Cost is a practical consideration too. Graphs reduce token burn. Instead of re-injecting full conversation histories or large document chunks on every call, an agent retrieves specific structured facts. Google’s Vertex AI offers context caching to cut costs on long-context production workloads, and a well-maintained graph amplifies that efficiency by making each retrieval tighter and more targeted.
AWS Context vs. Gemini Memory Bank: Two Different Bets
AWS Context is oriented toward existing enterprise data: the decades of accumulated records in data lakes and warehouses that most organisations have but cannot easily query with agents today. Its value is in making that dormant data reasoning-ready, with strong structural and semantic traversal for complex relational queries. The graph is largely pre-built from existing sources.
Gemini Enterprise’s Memory Bank takes the opposite angle: the graph grows from agent interactions. It is designed for systems that need to learn continuously from live use, personalised assistants, adaptive decision-makers, multi-agent workflows where operational parameters shift frequently. The ADK’s graph-based orchestration layer adds coordination across agent networks that AWS Context does not directly address.
Both move past the core limitation of simple vector search. The practical difference is whether a team’s primary need is structured access to pre-existing enterprise data (AWS Context) or a self-updating memory layer built from ongoing agent activity (Gemini Memory Bank). Many production deployments will eventually need elements of both.
Recommendations for Enterprise Adoption
Teams sitting on large volumes of structured and semi-structured enterprise data should look hard at AWS Context. The automated mapping promise is meaningful if data governance is already in reasonable shape: it converts accumulated institutional knowledge into a queryable asset agents can use immediately, with governed access and relational query support.
For teams building personalised or multi-agent systems where agents need to adapt from interactions over time, Gemini Enterprise’s Memory Bank is the more natural fit. The combination of dynamic memory, multimodal support and ADK orchestration covers the multi-agent coordination problem that static knowledge graphs handle less well.
Architecture selection here is worth taking seriously rather than defaulting to whatever the cloud vendor bundles, published benchmarks have found meaningful accuracy gaps between memory architectures on temporal queries. Agents that can reason over structured, relational, time-aware context consistently outperform those that cannot, and both AWS and Google have now shipped production tooling to close that gap. For more on AI agents and automation tools, visit our AI Agents section.
