Agent memory in 2026 — SurrealDB, Cognee, and the rest

In short

In 2026, AI agent memory is no longer a secondary concern: it is a market. SurrealDB 3.0 and Cognee raised $23M and $7.5M respectively within a few months, positioning context persistence as critical infrastructure. The classic five-database architectures — vectors, graph, relational, cache, files — are in the crosshairs. But only 27% of organizations have knowledge graphs in production. The sector is accelerating, but not yet mature.

SurrealDB 3.0: the database that wants to unify everything

On February 17, 2026, SurrealDB simultaneously announced a Series A extension to $23M and the launch of version 3.0. The round brings the total Series A to $38M and cumulative funding to $44M, with Chalfen Ventures and Begin Capital as new entrants alongside FirstMark and Georgian.

The positioning is explicit: “the future of AI agent memory”. SurrealDB moves away from the territory of general-purpose multi-model databases to claim the status of infrastructure platform for AI agents.

Context graphs first-class. The central new feature of version 3.0 is native integration of context graphs — graph structures that keep agent context synchronized as close to the data as possible, without external middleware. The declared objective is to reduce the glue code that characterizes current RAG architectures, where orchestration between vector, graph, and relational databases produces significant operational complexity.

Unification of vector + graph + SQL. SurrealDB 3.0 makes it possible to execute in a single SurrealQL query operations that previously required three distinct systems: vector search (top-k), graph traversal, and SQL filters. VentureBeat summarizes the positioning: “SurrealDB 3.0 wants to replace your five-database RAG stack with one.”

Surrealism. SurrealDB 3.0’s new WebAssembly extension system allows business logic and models to be executed directly inside the database. Combined with computed fields (logic defined once in the schema, evaluated at query time) and native file support (images, audio, documents queryable in SurrealQL), version 3.0 aims for maximum co-location between data and processing — a principle echoing the data-centric AI architectures advocated over recent years.

The conceptual distinction SurrealDB highlights — context graphs versus knowledge graphs — deserves attention. A knowledge graph stores durable semantic relationships (entity-relation-entity). A context graph maintains the state of the conversation and agent context in the short and medium term. SurrealDB is not replacing Neo4j on its own ground; it targets a different, more ephemeral and more operational layer.

In short: a knowledge graph is the fixed encyclopedia the agent consults. A context graph is the notebook it keeps during its session. Both coexist in a mature memory stack.

Cognee: open-source memory

Cognee presents itself as an open-source memory engine for AI agents. The core mechanic: ingest data from 38+ sources, structure it into a knowledge graph with embeddings and semantic relationships, and make it queryable via a unified interface. Three storage layers coexist: relational, vector, graph.

The ECL pipeline. Cognee’s central architecture rests on three phases: Extract (ingestion and parsing), Cognify (knowledge graph construction, semantic enrichment), Load (indexing and exposure for retrieval). This ECL pipeline is the core of the open-source project and its main technical contribution.

Memify Pipeline. Announced in 2025, Memify is a modular post-processing component that refines the knowledge graph through feedback loops. Evaluated responses feed the weights of the graph’s edges: memory becomes more precise with use, without full re-ingestion. This is the practical implementation of a principle often theorized — memory that learns from its mistakes.

Adoption. In 2025, Cognee went from ~2,000 pipeline executions to over one million — 500× growth in a year. The project is deployed in production at more than 70 companies, including Bayer for scientific research workflows. The GitHub repo exceeds 12,000 stars with 80+ contributors, and has earned the GitHub Secure Open Source Program certification.

Funding. In 2025, Cognee raised $7.5M in a seed round led by Pebblebed — Pamela Vagata (OpenAI co-founder) and Keith Adams (Facebook AI Research Lab founder) — with 42CAP and Vermilion Ventures, and angels from Google DeepMind, n8n, and Snowplow. The alignment of investors with AI infrastructure profiles is notable.

Where SurrealDB plays on the unification of storage layers, Cognee plays on the semantic depth of the graph and the continuous improvement loop. The two approaches are not mutually exclusive — several production stacks use Cognee on top of SurrealDB or Neo4j.

In short: SurrealDB plays the role of the floor — a unified store. Cognee plays the role of the smart shelves placed on top, which arrange and rearrange documents over time of usage.

The toolbox

The Graph-based Agent Memory survey (Yang et al., arXiv:2602.05665, February 2026) inventories the ecosystem of graph memory tools for agents. Among the most cited:

Mem0, Zep, Graphiti are the three most operationally ready open-source tools. Mem0 stands out with entity-centric retrieval that reduces noise by focusing on relevant nodes. Zep uses breadth-first expansion that explores successive neighbors of a target node. Graphiti implements the bi-temporal model: each fact is tagged with its validity time (when the event occurred) and its transaction time (when it was recorded), making it possible to resolve contradictions through temporal invalidation rather than overwriting — a critical property for long-horizon agents.

HyperGraphRAG represents a distinct research direction: hypergraphs connect an arbitrary number of nodes via hyperedges, preserving n-ary relationships that classical binary graphs fragment. The dual-retrieval of entities and associated hyperedges improves accuracy on complex queries involving multiple entities simultaneously.

2026 trends

The GNN → Graph Transformers shift. The most significant trend in graph memory architectures is the move from Graph Neural Networks (GNNs) to Graph Transformers. The primary motivation is hardware-based: Transformers are optimized for dense matrix operations on GPUs, whereas GNNs rely on sparse graph traversals that are less well supported by current hardware. Hybrid GNN+Transformer architectures (such as GIT-CD) outperform standalone modules on community detection and single-cell genomics, and are recommended for large-scale tasks or heterogeneous data.

In short: three converging signals — growing funding (~$50M cumulative over 2 years), accelerating adoption (Cognee × 500 in one year), hardware shift GNN → Transformers. The sector is moving from prototype to infrastructure, but 73% of companies still have no knowledge graph in production.

27% in production. At the end of 2025, only 27% of AI adopters reported having knowledge graphs in production, compared to 26% at the start of 2024 — a marginal increase over 18 months (source: Year of the Graph Newsletter, Autumn 2025). The gap between the discourse — omnipresent since 2023 — and actual adoption remains significant. The identified obstacles converge: operational complexity of graph maintenance, initial construction cost, lack of standardized tooling for incremental updates.

Context graph vs. knowledge graph. The defining debate of 2026 pits two philosophies against each other. Knowledge graphs (Cognee, Neo4j, Graphiti) structure durable semantic relationships: what the agent knows about the world, its entities, their attributes and connections. Context graphs (SurrealDB 3.0) maintain the operational state during a session: which tool was called, which decision was made, which intermediate result is available. It is not a binary choice — the most robust memory systems combine both levels — but the distinction clarifies why two players can both claim to do “graphs for agents” without directly overlapping.

Memory as infrastructure. The convergence of funding rounds (SurrealDB, Cognee, Mem0 which preceded them in 2024) signals a shift in perception: memory persistence is moving from the status of application feature to that of infrastructure. Teams building long-lived multi-agent systems can no longer treat memory as a secondary problem to be solved with Redis and a few embeddings.

What to remember

• SurrealDB 3.0 ($44M raised) unifies vector, graph, and SQL in a single query and integrates context graphs natively — position: replace multi-database RAG stacks.
• Cognee ($7.5M, 12K stars, 70 companies) offers an open-source ECL pipeline with Memify feedback loop — position: semantic layer on top of any storage backend.
• Graphiti, Mem0, Zep are the three most widely used operational open-source tools; Graphiti stands out for its bi-temporal model, essential for long-horizon agents.
• 27% production adoption of knowledge graphs at end of 2025: the acceleration of discourse has not yet produced adoption acceleration at the same pace.
• The GNN → Graph Transformers shift is driven by hardware constraints, not just performance arguments — Transformers win because the hardware favors them.