Our KG Memory Transfer paper has been accepted at RLC 2026
Our paper Short-Term-to-Long-Term Memory Transfer for Knowledge Graphs under Partial Observability has been accepted at RLC 2026, the Reinforcement Learning Conference.
The paper is also now part of the broader KG Memory Transfer research page, where we explain the method, results, and implementation context in more detail.
The question
This work studies a specific memory problem inside partially observable reinforcement learning: when short-term symbolic observations arrive under memory constraints, which facts should be transferred into long-term memory and which should be dropped?
Our approach treats that decision as an explicit keep/drop policy over knowledge-graph facts. Rather than hiding consolidation inside a latent recurrent state, the method keeps transfer decisions inspectable at the level of individual triples. That makes the learned behavior easier to analyze while still improving downstream question answering in the RoomKG setting.
What it shows
The paper treats the move from short-term observation to long-term memory as a learned decision problem. It shows that selective symbolic transfer can outperform both symbolic heuristics and history-based neural baselines in a controlled setting.
Resources