Why AI agents need a multi-model foundation

Q: What is the context layer?

The context layer is a unified database infrastructure that combines storage, context, and memory in a single ACID transaction. It replaces the fragmented stack of document stores, graph databases, vector databases, and memory systems with one engine.

Q: How does the context layer differ from knowledge graphs?

Knowledge graphs represent stable, ontological facts - what an organisation knows. The context layer adds dynamic, situational relevance - what is happening right now, and why it matters. A well-designed context graph anchors itself to the canonical facts of a knowledge graph while capturing the lived experience of AI agents.

Q: Why can't memory middleware solve the context problem?

Memory middleware sits above the storage layer rather than within it. It abstracts the fragmented infrastructure but doesn't eliminate it. When context changes in one underlying system, synchronisation to others is not guaranteed. The semantic drift problem persists underneath the abstraction.

Beyond knowledge graphs to real-time contextual reasoning. The context layer must live in the database, not above it.

Explore the platform Download PDF

EXECUTIVE SUMMARY

The context wall

Enterprise AI initiatives share a common failure pattern. The cause is not the model - it is the architecture beneath the model. As LLMs mature into autonomous agents, enterprises are hitting the Context Wall.

The problem

LLMs are stateless. RAG is a temporary patch. Agents need living memory. Current architectures rely on fragmented frankenstacks - disparate vector, graph, and document databases stitched together with glue code.

The thesis

The bottleneck is architectural, not model-related. Single-model stores and fragmented data platforms are structurally misaligned with the requirements of agentic state.

The solution

A unified Context Graph built on a multi-model foundation that treats storage as an active participant in reasoning - not a passive retrieval layer.

THE CONTEXT WALL

Why agents fail in production

Nearly every organisation attempted to deploy agentic workflows in 2024 and 2025. Most failed - not because models are incapable, but because agents lack persistent, structured, real-time context.

Failed pilots

Agents produced wrong answers - not because the model couldn't reason, but because it lacked context for how terms were defined, which tables were authoritative, and how definitions had changed.

Cost and governance

Running production agents on four specialist databases means four sets of infrastructure, four sync pipelines, four access control systems, and four failure modes to debug.

Architecture, not models

Longer context windows don't solve the fundamental issue. An agent that cannot access reliable, structured, real-time context will hallucinate regardless of model capability.

THE READ-THINK-WRITE LOOP

Why AI needs a context layer

Agents operate through a continuous cycle: perceive the environment, reason over memory, commit an action back to storage. This loop runs in milliseconds and never stops. Analytical platforms are not built for this.

READ

Graph traversal, vector search, and temporal facts - one SurrealQL statement, one round trip.

SurrealQL

Multi-model query

Graph

Vector

Temporal

Unified result set

THINK

Documents, relationships, embeddings, and history arrive together. The agent reasons over complete context.

Result set

Multi-model query output

Documents

Relations

Vectors

History

Complete context

WRITE

Persist decisions, update entities, trigger events - one ACID transaction, no partial writes.

BEGIN TRANSACTION

ACID transaction boundary

Update

Create

Trigger

COMMITTED

THE FRAGMENTED MEMORY TAX

Why multi-model wins

When an agent's memory is split across multiple systems, unification happens at the application layer. This creates latency, semantic drift, loss of dimensionality, and consistency gaps that compound under load.

App state

User session

Vector DB

Embeddings

Graph DB

Relations

Doc store

Documents

Auth

Identity

Data fragments at every system boundary

Agent

Vector DB

Graph DB

Doc store

Auth

Total

Cumulative latency

Latency of reconstruction

Every memory retrieval requires multiple async calls to different APIs, joined at the application layer. Milliseconds compound into seconds.

Semantic drift

When a document is updated but its vector embedding or graph relationships aren't refreshed, the agent reasons over contradictory information.

Loss of dimensionality

Vector search finds similarity, but can't explain why - the relationship - or show the constraints without a multi-stage join.

Consistency boundaries

Each system has its own consistency model. The agent's view depends on which system it queries and when.

SURREALQL IN ACTION

Implementing semantic context

SurrealQL collapses complex multi-system operations into single, atomic queries. Graph traversal, vector search, and document operations in one statement.

RICH EDGES

Graph + document + vector in one operation.

RELATE agent:researcher  ->interacts_with  ->user:customer123SET  time = time::now(),  topic = "Billing Dispute",  sentiment_score = 0.2,  context_embedding =    fn::embed("frustrated"),  summary = "Double-billing    on invoice #55";

UNIFIED SEARCH

Vector similarity + graph filtering in one query.

SELECT * FROM interacts_withWHERE context_embedding  <|2|> fn::embed(    "billing issues"  )AND out.purchased_product  = product:enterprise_saasAND time  > time::now() - 30d;

LIVE QUERIES

Multi-agent sync without polling.

LIVE SELECT * FROM taskWHERE status = 'Completed'AND project  = project:ai_migration;

THE ARCHITECTURE

Enterprise Semantic Foundation

The ESF is the architectural response to semantic drift. A unified, multi-model substrate that serves as the common language for the entire organisation - anchoring canonical meaning, activating real-time context, and governing access in one system.

Applications

Spectron

Memory

Entity extraction

Knowledge graph

Temporal facts

Hybrid retrieval

SurrealDB

Context

Documents

Graphs

Vectors

Time-series

Auth

APIs

Storage

Quorum consensus

Compute-storage sep.

Scale to zero

Object storage (S3 / GCS / Azure Blob)

Anchoring canonical meaning

The Knowledge Graph houses high-integrity, governed facts - the gold-standard data that has passed through traditional ETL and quality checks.

Activating real-time meaning

The Context Graph captures the high-velocity, lived experience of AI agents - always anchored to canonical truth.

Unified access control

Permissions are defined once at the data layer. Whether vector search, graph traversal, or document lookup - same security constraints.

Explainable AI

Because relationships are first-class citizens, the path an agent took to reach a conclusion is traceable through the graph.

TRUSTED BY

Enterprise teams building on SurrealDB

From knowledge graphs to AI assistants - how enterprise teams are building on the context layer.

Samsung

Unlocking insights with knowledge graphs

Samsung Ads uses SurrealDB to build dynamic, real-time knowledge graphs for smarter campaign execution - collapsing three legacy data stores into one.

Read case study

Verizon

AI assistant empowering 10,000 technicians

Verizon uses SurrealDB to power a generative AI assistant for 10,000 field technicians, delivering instant access to documentation, outage updates, and workflows.

Read case study

Tencent