RecallAI

A team of four—Aman Kumar, Rishu, Abhishek and Ayan—participated in the Graph Hacks hackathon co-hosted by WeMakeDevs and FalkorDB, chose the 'Agent Memory and Coordination' track, and built the RecallAI prototype in four days.

RecallAI's architecture consists of four main parts: a frontend chat interface, backend request handling and memory retrieval coordination, a language model, and the FalkorDB graph database.

The team chose FalkorDB to explore a graph-based memory structure rather than storing each conversation as an isolated text chunk.

The team chose Qwen2.5-1.5B-Instruct as the language model because the hackathon schedule was tight and hardware was limited, training a model from scratch was unrealistic, and they wanted to focus their effort on the memory architecture.

Testing on day three found that the system sometimes recalled information when it could recognize keywords or quote directly, but results were unstable when questions depended on broader conversational context, sometimes behaving as if it had no memory or giving confident answers that did not reflect the earlier conversation.

The possible causes of memory system failure are distributed across five stages: whether information was correctly saved, whether the correct information was retrieved from FalkorDB, whether the backend passed the retrieved information to the model, whether the model had enough context to understand the associations, and whether the model ignored or misread the information it received.

On day four the team decided not to add new features and instead focused their effort on six core behaviors: saving useful information, retrieving the correct memories, passing information to the model, using context correctly, avoiding irrelevant memories, and isolating user memories.