BlazorMemory 1.0 launches with local AI memory for .NET apps

A year after starting the project in January, developer Aftab Khan has released BlazorMemory 1.0, a .NET library designed to give AI chat assistants in Blazor WebAssembly (WASM) applications the ability to remember context without relying on external servers or vector databases. Khan began the project after discovering that existing solutions were primarily built for Python environments and assumed the presence of a server infrastructure, which Blazor WASM applications lack due to their client-side execution model.

The 1.0 release includes four storage backends to accommodate different deployment scenarios. IndexedDB enables pure client-side storage with no backend required, making it ideal for standalone Blazor WASM applications. For server-backed applications, the library supports EF Core for SQL Server and SQLite, and pgvector for PostgreSQL with native vector search capabilities using HNSW indexing. An InMemory backend is included for testing purposes, ensuring developers can prototype without additional dependencies.

BlazorMemory also integrates with four AI providers for embeddings and fact extraction: OpenAI, Anthropic, Azure OpenAI, and Ollama. These providers can be mixed and matched, allowing developers to use Ollama for embeddings while leveraging Anthropic for fact extraction, for example. The Ollama integration stands out for its ability to operate entirely locally, eliminating the need for API keys or per-request costs. This makes it particularly attractive for development and prototyping, as it removes the last barrier to experimenting with memory-enabled AI assistants in Blazor.

The library ships with a drop-in UI component that provides a prebuilt interface for managing memories. The component includes features such as viewing stored memories, deleting or clearing memories, exporting and importing memory data, and providing feedback via thumbs up or down buttons. Additionally, a mode toggle allows users to switch between different operational states. The UI is implemented in vanilla JavaScript with a force-directed graph visualization for memories, where nodes represent individual memories and edges connect memories based on cosine similarity thresholds.

BlazorMemory also integrates with Semantic Kernel, a Microsoft framework for building AI agents. By implementing Semantic Kernel's IMemoryStore interface, BlazorMemory can be dropped into existing Semantic Kernel applications without requiring any additional changes to the codebase. This compatibility broadens the library's appeal to developers already invested in the Semantic Kernel ecosystem.

One of the project's most notable features is its support for multi-agent shared memory. Multiple AI agents can share a single memory pool, with each agent writing to its own namespace while reading across all namespaces. This enables agents to learn from each other's experiences, fostering collaborative behavior within applications. The implementation required careful design to avoid redundancy and ensure efficient memory consolidation.

The development process was not without challenges, particularly around the library's storage interface, which underwent three major revisions. Early iterations used positional parameters, which led to breaking changes whenever new features like namespaces or importance scores were added. Khan noted that adopting an options object pattern from the start would have made these changes additive rather than disruptive. Consolidation of memories also proved more complex than initially anticipated, as early versions frequently stored near-duplicate entries such as "User works in software," "User is a developer," and "User writes code."

To address this, Khan implemented a priority order in the consolidation prompt, instructing the model to first consider not adding a new memory, then updating existing ones, deleting irrelevant entries, and finally adding new memories only if justified. This single change significantly improved memory quality and reduced redundancy. Another hurdle involved parsing JSON responses from local models, which often included markdown fences or extraneous commentary. The solution involved stripping code fences, locating the first bracket, and retrying with a stricter prompt if parsing failed.

Khan emphasized the importance of shipping small, frequent releases. Despite the library consisting of 14 packages, each release was deliberately incremental, with features like namespaces, export/import functionality, and feedback mechanisms added in separate versions. This approach not only made development manageable but also generated consistent engagement, as downloads spiked following blog posts rather than standalone releases. The project is now MIT licensed, includes 132 tests, and is ready for production use.

For developers building AI-powered applications in Blazor, BlazorMemory 1.0 offers a compelling solution for adding memory capabilities without the overhead of server infrastructure. Its integration with multiple AI providers, support for local operation, and compatibility with Semantic Kernel make it a versatile tool for both prototyping and production environments. The library's focus on reducing redundancy and improving memory quality through iterative refinement demonstrates the importance of addressing real-world usage challenges in AI development.