Category #1: AI Databases & Retrieval

What is this category about?

• The retrieval layer for AI applications. These companies provide the databases and search infrastructure that allow AI applications to store large volumes of information and retrieve the most relevant pieces when they are needed.

• Making unstructured data searchable. Traditional databases work well when information follows a fixed structure. AI applications often need to search across documents, images, code or other unstructured content based on meaning rather than exact keywords.

• Retrieving the right context quickly. An AI model cannot process an entire company database every time it answers a question. These systems identify the small amount of relevant information that should be added to the model’s context at runtime.

• Running retrieval reliably at scale. Retrieval becomes harder as the volume of data and number of users grow. These products help maintain low latency, control infrastructure costs and support production requirements such as security, access controls and data residency.

What do products in this category do?

• Startups in this category usually complement existing databases rather than replace them. A company’s operational data may remain in systems such as PostgreSQL or a data warehouse. The retrieval layer creates a searchable representation of that information so an AI application can find the right context quickly.

• The retrieval layer is usually built as a synchronized index. It does not usually compile or replace the main database. Instead, data is copied from existing databases, transformed into searchable text and embeddings, then indexed in the retrieval layer. The index is updated through scheduled syncs or real-time events when the underlying data changes. The main complexity comes from keeping this layer “fresh”, secure and consistent with the source.

• Many products rely on vector search. A vector database converts content into numerical representations called embeddings. This allows an application to find information with a similar meaning, even when the query uses different words from the source material.

• Specialized vector databases focus on fast semantic retrieval. Companies such as Qdrant, Chroma and Pinecone provide infrastructure designed specifically for storing embeddings and searching across them at low latency. Their managed products increasingly add the security and operational controls required for production deployments.

• Search platforms combine semantic and keyword retrieval. Vespa, Meilisearch and turbopuffer support hybrid search, where vector similarity is combined with traditional full-text search and ranking. This approach is useful when exact terms remain important alongside broader semantic meaning.

• Some databases combine the application database and retrieval layer. Instead of storing normal application data in one database and embeddings in a separate vector database, SurrealDB (for example) can store both in the same system. It supports relational records, documents, graphs and vectors, allowing an application to manage its core data and search it semantically without constantly synchronizing several databases.

• Multimodal platforms extend retrieval beyond text. LanceDB is built around data such as images, video and embeddings, while also supporting workflows related to dataset preparation and model training. Its approach is closer to an AI-focused lakehouse than a standalone vector database.

• The products are gradually moving beyond basic retrieval. Weaviate has added agent services and managed memory, while Pinecone is expanding into knowledge infrastructure with connectors and agent-facing context tools. Chroma has also introduced managed ingestion that continuously turns repositories and websites into searchable data.

What are the major trends shaping this category?

• Specialized vendors increasingly compete with databases companies already use. Of course traditional databases and search engines are adding vector and hybrid search directly to their existing products. This allows teams to keep application data and embeddings in the same system, which may be sufficient for simpler AI applications and avoids operating a separate retrieval database.

• Hyperscalers have a strong distribution advantage. AWS, Google Cloud and Microsoft can bundle retrieval into the broader cloud environments where enterprise data already lives. Their advantage is not necessarily a better retrieval engine, but easier procurement, existing security controls and fewer new infrastructure vendors for customers to manage.

• The category is likely to consolidate around broader platforms. Vector search is becoming a feature inside databases, search engines, lakehouses and agent infrastructure. This suggests that standalone retrieval products will increasingly need to expand into a larger data or context platform, or become deeply differentiated around a specific scale or workload.

How does the funding environment look?

• The category has attracted substantial capital, but funding is concentrated among a few leaders. Pinecone raised a $100 million Series B, while Weaviate, Zilliz and Qdrant each secured rounds of $50 million or more. This reflects the large infrastructure opportunity investors see around AI retrieval, but also the capital required to build and operate these platforms at scale.

• Funding remains available for companies with a differentiated technical position. Qdrant raised a $50 million Series B in March 2026, SurrealDB secured a $23 million Series A extension in February 2026 and LanceDB raised a $30 million Series A in 2025. These rounds suggest investors are still willing to finance new challengers rather than treating the market as already settled.

• Capital is increasingly backing broader platform ambitions. Recent funding has gone to companies expanding beyond a standalone vector database. SurrealDB is positioning itself as an application database and agent memory layer, while LanceDB is building a multimodal data platform that also supports data preparation and model training.

• The market appears to have moved beyond its initial funding wave. Several prominent companies last raised major rounds in 2022 or 2023, including Pinecone, Weaviate, Zilliz, Meilisearch and Chroma. The next phase will therefore depend less on excitement around vector databases and more on whether these companies can convert adoption into durable cloud and enterprise revenue.

Category #2: Memory & Context Infrastructure

What is this category about?

• This is the layer that gives AI applications continuity. These products help AI applications retain useful information across conversations and actions. Instead of treating every interaction as a new session, an agent can remember what it has learned about a user, a task or an organization.

• Preventing agents from forgetting between sessions. Language models do not naturally preserve information after a conversation ends. A memory layer stores the relevant parts of previous interactions and makes them available when the agent is used again.

• Keeping the model’s context focused. Sending an entire conversation history to a model quickly becomes expensive and unreliable. Memory systems select the information that matters for the current task and leave out details that would distract the model.

• Handling information that changes over time. A user’s preferences, a project’s status or a company’s data can evolve. These products help agents update old memories, recognize when facts have changed and prioritize recent information when appropriate.

• Sharing context across agents and interfaces. A user may interact with the same AI system through a chatbot, coding tool or business application. A shared memory layer allows these different interfaces to access the same context rather than developing separate histories.

What do products in this category do?

• They sit between the AI application and the model. The memory layer observes interactions, stores selected information and retrieves the most relevant memories before each model call.

• They usually sit above the retrieval layer. The retrieval system finds relevant information in documents or databases, while the memory system decides what an agent should retain from previous interactions and how that knowledge should change over time. Memory products often use a vector database or another retrieval engine underneath, but add the logic for remembering, updating and forgetting information.

• Memory can be stored in several forms. Some information remains as summaries or individual facts, while more complex systems represent relationships between people, events and entities in a knowledge graph. The system can then combine these forms when building the context sent to the model.

• Memory as an external API: The memory system sits alongside the agent and automatically extracts, stores and retrieves useful information from previous interactions. Mem0 follows this approach, giving developers a dedicated memory service without requiring them to build the underlying storage and memory logic themselves.

• Memory as a temporal knowledge graph: Information is stored as connected facts, entities and events, while preserving how these relationships change over time. Zep uses this approach to distinguish current information from outdated facts and assemble a compact context for each interaction.

• Memory built from documents and interactions: These systems combine conversation history with information from files, business tools and other data sources to create a shared semantic memory. Cognee illustrates this approach by turning different types of data into searchable and connected knowledge that agents can reuse.

• Memory integrated into the agent runtime: Instead of adding memory as a separate service, the agent framework manages memory as part of how the agent operates. Letta follows this model, allowing agents to inspect, update and reorganize their own persistent context through memory blocks and context repositories.

• Memory as an organizational context layer: This approach continuously collects information and turns it into shared context that several agents or applications can access. Graphlit pursued this model through connectors, extracted facts and relationships, before moving toward more specific applications and eventually winding down its original platform.

What are the major trends shaping this category?

• Memory is becoming part of a broader context engineering layer. The central challenge is no longer only storing previous messages. Products increasingly decide what information the model should receive, how it should be formatted and which sources deserve the most weight.

• Larger context windows do not remove the need for memory. Models can process more information than before, but filling the context window with every past interaction remains costly and often reduces answer quality. Memory infrastructure is therefore shifting from overcoming a hard technical limit to improving context quality and efficiency.

• The category is converging with retrieval infrastructure. Memory systems use vector search and knowledge graphs, while retrieval databases are adding persistent memory and context services. The boundary between the two categories is becoming less clear as vendors expand across the stack.

• Memory is also converging with agent runtimes. Letta has moved from a standalone memory platform toward a complete agent harness, while other vendors are adding tools that control how agents learn and manage context. This suggests memory may become a core feature of agent platforms rather than remain a separate product category.

• Enterprise adoption requires governed memory. Persistent context can contain sensitive information and may influence many future agent actions. Products are therefore adding stronger access controls, source tracing and private deployment options so companies can inspect where a memory came from and control who can use it.

• The market remains highly technical and research driven. Cognee, Mem0, Zep and Letta are actively hiring for platform engineering, applied AI and memory research roles. This indicates that reliable long term memory remains an open technical problem rather than a standardized infrastructure component.

How does the funding environment look?

• The category is funded, but remains much earlier than the retrieval database market. The category on the map currently records approximately $47 million in disclosed financing across these five companies. Most rounds are still at the seed stage, reflecting a market where the core technology and commercial model are both developing.

• Mem0 has attracted the strongest funding momentum. The company raised a $20 million Series A in October 2025 after an earlier $3.9 million seed round. It is the only company in this category with a recorded Series A, making it the clearest current funding leader.

• Investors are still backing new technical approaches. Cognee raised a $7.5 million seed round in February 2026, while Letta launched with a $10 million seed round in 2024. These investments show continued interest in both independent memory layers and broader stateful-agent platforms.

• Graphlit’s wind down highlights the category’s commercial risk. Graphlit raised a $3.56 million seed round in 2025 but closed new signups for its original service in July 2026..

• Acquisition activity has not started yet. The Axomap map contains no acquisitions involving companies in this category. The market remains in an experimentation phase, although its overlap with databases, agent platforms and model providers could eventually create several paths toward consolidation.

Category #3: Data Ingestion & Transformation

What is this category about?

• The layer that turns raw enterprise data into usable AI context. These products collect information from documents and business systems, transform it into a format AI applications can understand and keep it available for downstream retrieval or agent workflows.

• Making difficult documents machine-readable. Important enterprise information is often trapped in PDFs, presentations, scans, spreadsheets and forms. These files contain layouts, tables and visual relationships that cannot be reliably understood through basic text extraction alone.

• Converting unstructured content into consistent outputs. AI applications work better when information follows a predictable structure. These products identify the relevant fields, sections and relationships inside a document, then convert them into structured records, searchable chunks or validated JSON.

• Keeping AI systems connected to changing data. Enterprise information is continuously updated across drives, databases and operational applications. Ingestion platforms synchronize these sources so retrieval systems do not rely on a static and increasingly outdated copy of the data.

What do products in this category do?

• They generally sit above the retrieval and memory layers. Ingestion products prepare and synchronize the information, retrieval systems search it and memory systems decide what an agent should retain over time. Some platforms cover several of these steps, but the core role of this category is to turn raw data into context that the two other layers can use.

• They do not usually replace the source origin. Documents remain in tools such as Google Drive, SharePoint or an object store, while operational records remain in the company’s databases. The ingestion layer reads from these sources, processes the content and updates the AI-ready version when the original information changes.

• Continuous ingestion keeps context synchronized. These products connect to enterprise systems and detect new or modified data through scheduled syncs, event notifications or live data streams. Pathway specializes in continuously processing changing data, while Ragie combines managed connectors with automatic synchronization for AI applications.

• Parsing infrastructure prepares documents for retrieval. This approach identifies headings, tables, images and reading order before breaking a document into meaningful sections. Unstructured and LlamaIndex provide parsing and ingestion tools designed to produce better inputs for search, RAG and agent systems.

• Document extraction turns files into structured records. Instead of preparing documents only for search, these products extract specific fields into a defined schema that can feed databases and business workflows. Reducto and Anyformat use this approach for complex enterprise documents, while Cradl AI focuses on configurable document-automation workflows.

• Validation first platforms combine extraction with operational controls. The system checks whether extracted values are internally consistent, links them to their original location and sends uncertain cases for human review. Holofin applies this approach to financial and administrative documents, where auditability and reconciliation are essential.

• Visual workflow products make pipelines easier to configure. Document processing often requires several connected steps rather than a single model call. Anyformat, Reducto, Unstructured and Holofin provide interfaces for combining classification, parsing, extraction, validation and export into a reusable workflow.

• Broader context platforms extend beyond ingestion. Contextual AI combines document processing with retrieval, ranking, generation and agent orchestration. This illustrates how some vendors are trying to own the complete path from proprietary enterprise data to a grounded AI application.

What are the major trends shaping this category?

• Document intelligence is moving beyond traditional OCR. OCR can recognize characters, but it does not reliably understand document structure or the relationships between fields. Modern products combine visual models, layout analysis and language models inside a broader pipeline to interpret complex documents.

• Better multimodal models are expanding the market but also increasing competition. Improvements in foundation models (OpenAI, Claude…) make it easier to understand tables, scans and images. This lowers the technical barrier for new vendors, while pushing specialized companies to differentiate through reliability, workflow control and enterprise deployment rather than basic extraction alone.

• The simplest document use cases are becoming commoditized. Standard invoices and clearly formatted forms can increasingly be processed by cloud platforms, automation tools or general-purpose models. Specialized vendors are therefore moving toward longer documents, irregular layouts and regulated workflows where generic tools still struggle.

• Reliability is becoming more important than raw extraction accuracy. Anyformat, Reducto and Holofin are emphasizing benchmarks, confidence scores and source linked outputs, while Cradl AI uses model cross checking and human review. The market is moving toward systems that can explain and verify an output rather than simply produce one.

• Ingestion and retrieval are converging into broader context platforms. Ragie already combines parsing, synchronization and retrieval, while Unstructured and LlamaIndex have expanded beyond basic document processing into managed AI data workflows. Buyers may increasingly prefer one platform that moves data from the source system to the agent.

• The category is moving closer to business workflows. Extraction is increasingly followed by classification, fraud detection, validation, routing and approval. This creates more value than parsing alone, but also moves vendors into competition with document automation and vertical software products.

• Regulated industries are becoming an important commercial wedge. Unstructured is expanding into government and defense, while Reducto has highlighted healthcare and financial-services use cases. Holofin focuses directly on financial documents, showing how trust requirements can create defensible opportunities for more specialized platforms.

How does the funding environment look?

• This is one of the most heavily funded parts of our data stack. The companies in the landscape have raised more than $320 million in disclosed funding. 

• Reducto has emerged as the category’s strongest recent funding story. The company raised a $24.5 million Series A in April 2025 and a $75 million Series B only six months later. 

• Contextual AI attracted major funding around a broader platform thesis. The company raised a $20 million seed round in 2023 and an $80 million Series A in 2024. Its funding was based on an ambition to combine ingestion, retrieval and grounded generation rather than sell a standalone parsing product.

• Unstructured has also raised significant capital. The company secured $25 million across its seed and Series A financing, followed by a $40 million Series B in 2024. Its investors include major data and infrastructure companies, which reflects the strategic importance of preparing unstructured enterprise data for AI.

• Acquisition and consolidation activity has started. Reducto announced the acquisition of Opennote in May 2026, showing that the best-funded vendors are beginning to acquire adjacent technology and talent.

• Strategic talent deals are another possible consolidation path. In May 2026, Google DeepMind reached a licensing and hiring deal that brought more than 20 Contextual AI researchers, including its co-founder and CEO, into Google. The company was not acquired outright, but the deal shows how hyperscalers can absorb valuable teams and technology without purchasing the full business.

• The category is attractive but increasingly demanding. Investors clearly believe that proprietary enterprise data will require a dedicated AI preparation layer. At the same time, vendors must compete with foundation model improvements, cloud platforms and enterprise data incumbents, making technical differentiation and distribution increasingly important.