EgoCITE gives AI agents searchable first-person memory at 36x lower cost
New framework improves egocentric QA accuracy by up to 14.2% using time-aware retrieval.
EgoCITE (Egocentric Context-augmented Indexing and Time-aware Evidence retrieval) is a new framework from researchers Le Zhang and Ke Sun, posted on arXiv, designed to solve two core bottlenecks in egocentric memory systems. First, existing indices built from context-poor captions are unreliable for agentic search. Second, retrieval systems ignore the temporal intent embedded in questions. EgoCITE addresses both with three components: EgoScheme uses local multimodal context to transform fragmentary video captions and speech transcripts into self-contained atomic memory indices; EgoIndex organizes complementary action, activity, utterance, and conversation representations into multi-view memory indices at multiple granularities; and EgoRetrv combines semantic search with question-conditioned temporal relevance scoring and curated evidence retrieval.
The framework was evaluated on three egocentric QA benchmarks—EgoLifeQA, EgoMem, and EgoR1-Bench—measuring answer accuracy and target-event retrieval alignment. Results show EgoCITE improves accuracy over agentic memory baselines by at least 4.4% and up to 14.2%, while achieving 36x lower cost compared to long-context LLM agents. This suggests that building structured, time-aware memory indices is far more efficient than shoving entire lifelogs into a model's context window. For wearable AI, AR glasses, and personal assistants, this approach makes long-horizon memory practical: AI can recall what you did, said, or saw hours or days ago with precise temporal grounding, without paying massive context-window bills.
- EgoCITE improves egocentric QA accuracy by 4.4-14.2% over agentic memory baselines
- The framework uses time-aware retrieval to respect the temporal intent of questions
- Achieves 36x lower cost than long-context LLM agents on egocentric memory tasks
Why It Matters
EgoCITE makes long-horizon personal memory practical for agents and wearables, enabling efficient, time-aware recall without costly long-context inference.