The problem
An agent's memory needs more than a relevant search result. Facts can change, and a retrieved claim needs a source that someone can inspect. Keeping that information local is another constraint.
TelloDB explores how to handle retrieval, fact history, and source tracking in one memory engine.
How the prototype addresses it
The prototype runs in one Rust process and uses a separate SQLite database for each tenant. It combines keyword and vector search, graph relations, and local inference.
Past versions of facts remain available, and source tracking connects retrieved information to its origin. This makes history and provenance part of the memory model.
What the evaluation showed
In one recorded LoCoMo run, an annotated source session appeared among the top eight results for 95.8% of eligible questions.
That measures retrieval of a source session. It does not mean that 95.8% of generated answers were correct, or establish the same result on other datasets. The paper describes the evaluation and its limits.
Implementation
Rust · SQLite / FTS5 · HNSW · ONNX