In-house product · Personal AI
Mementri connects thoughts with their original context.
Mementri is a personal AI memory for thoughts, tasks, ideas, projects and recollections. People can capture an entry by voice or text. Several specialised processing stages then organise and connect it while preserving the original input.
Product concept, system architecture, development, AI infrastructure and ongoing operation by ProWebSolutions.
Turning speech and text into a traceable memory
The starting point is intentionally simple: a thought is captured by voice or text. Spoken entries pass through self-operated, Whisper-based speech-to-text infrastructure. The original transcript remains available. On top of that source, Mementri can identify whether an entry contains a task, an idea, a project reference or a question.
Keeping the source separate from derived information is a central product principle. A later summary or classification does not replace what was actually said or written. Users can therefore return to the original wording and review the processed information in its proper context.
Specialised stages instead of one monolithic prompt
Mementri does not process an entry through one large AI request. Multiple specialised stages handle clearly defined responsibilities. An early classification identifies relevant content types before deeper processing examines relationships, tasks, ideas or project context.
This architecture narrows the context for each decision and makes individual stages easier to examine. It also supports asynchronous execution: a broker, workers and queues distribute longer-running processing rather than tying it to a single web request.
Connections remain anchored to the source
Derived information is linked to precise passages in the original entry. Positions and cryptographic verification data help preserve that relationship. The system can therefore show not only a result, but also the source passage on which it is based.
When related entries may belong to the same thread, the candidate set is narrowed technically before the AI decides whether the connection is meaningful. Temporal consistency is considered as well, so relationships are not created solely because two entries contain similar terms.
Self-operated AI infrastructure in Frankfurt
Mementri uses a self-hosted language model with around 80 billion parameters in infrastructure located in Frankfurt. Speech recognition is also operated under direct control. Personal thoughts therefore do not have to be sent to an external commercial AI provider for the core processing workflow.
For ProWebSolutions, the project combines product development with AI operations. The web application, data model, linking logic, speech processing and compute-intensive background jobs have to work as one observable system. The decisive capability is not a single model, but the controlled chain from original input to a useful, connected memory.
Relevant engineering services
The capabilities demonstrated by this project.
Related current work
Explore related ProWebSolutions products.
In-house product · Language & audio
Verbison
Editorially controlled multilingual website audio with translation, speech synthesis and synchronised playback.
- Role
- Product concept, website integration, translation and text-to-speech pipeline, and ongoing operation.
In-house product · Browser monitoring
StoreVigil
Browser-based monitoring for websites and stores with visual issue location, repeated checks and distributed workers.
- Role
- Product concept, browser automation, monitoring architecture and ongoing operation.
In-house product · Geodata & routing
RouMeGo
Personalised walking, hiking and cycling routes based on OpenStreetMap, custom scoring and individual preferences.
- Role
- Product concept, geodata processing, routing, web application and native Android app.