Recuerdos.local: An Offline Ofrenda & Memory Weaver

This submission introduces Recuerdos.local, an offline digital ofrenda and memory archivist. The project is a fully local, voice-interactive tool that conducts gentle, conversational interviews to gather stories about lost loved ones. It preserves these oral histories without exposing them to corporate data mining.

Recuerdos.local was inspired by the tradition of Día de los Muertos, where families spend time reminiscing about departed relatives. However, these stories often fade over time. The goal was to create a respectful and secure way to archive these memories.

The project is built entirely on open-source AI components. Voice transcription is handled by Whisper.cpp, which supports both English and Spanish. The conversational logic is provided by Meta's Llama-3.1-8B-Instruct model, executed entirely on the user's local hardware via Ollama. A lightweight LangChain framework manages the state and guides the conversation, asking follow-up questions to elicit more detail.

Once the interview session concludes, the LangChain agent compiles the gathered information into a formatted Markdown file representing the digital ofrenda. This approach emphasizes privacy and data sovereignty. By keeping all processing local, the project ensures that intimate family stories remain within the home and are not used for any other purpose.

The significance of open innovation becomes apparent in the context of a sensitive topic like death and memory. When dealing with such personal and emotional content, privacy is not just an added feature but a fundamental requirement of respect. Running the project locally guarantees total data sovereignty—no data leaves the user's machine.

The design of Recuerdos.local is offline by nature. It can function perfectly without an internet connection, fitting the holiday theme of returning to the past. The project also offers zero ongoing costs, as it relies solely on local inference. This means it can be set up on an old laptop and left running during the holiday season without concerns about API billing.

A notable aspect of the project is the agent session, which demonstrates the evolution of prompt engineering and the seamless integration of Whisper.cpp and Llama-3.1. The interaction between the user and the agent reveals how the system can subtly prompt for additional details, capturing nuances that might otherwise be missed. This led to a poignant moment when the system remembered a specific Tejano radio station from a family member's past, preserving that detail permanently in the digital ofrenda.