Neuro-Symbolic Memory Transfer for Knowledge Graphs Under Partial Observability
A novel reinforcement learning technique has been developed to simulate the transfer of memory from short-term to long-term within knowledge graphs, even when faced with partial observability. This approach employs a neuro-symbolic, value-based decision-making process, allowing an agent to decide whether to retain or discard observed triples prior to their long-term integration. Utilizing a per-item Q-learning framework with shared parameters and temporal-difference updates, it effectively manages short-term buffers of varying sizes. In tests on the RoomKG benchmark, with a long-term memory capacity of 128, the learned transfer decisions surpassed both symbolic and neural benchmarks, including those utilizing temporal annotations and LSTM/Transformer models. The findings were published on arXiv with the identifier 2605.22142.
Key facts
- Reinforcement learning under partial observability requires deciding what information to retain.
- The study casts short-term-to-long-term transfer as a neuro-symbolic value-based decision problem.
- For each observed triple, the agent chooses to keep or drop it before long-term insertion.
- A per-item Q-learning design with shared parameters is used for variable-sized short-term buffers.
- Temporal-difference updates are applied over matched items across consecutive steps.
- RoomKG benchmark with long-term memory capacity 128 is used for evaluation.
- Learned transfer decisions outperform symbolic and neural baselines.
- Baselines include symbolic baselines with temporal annotations and history-based LSTM/Transformer baselines.
Entities
Institutions
- arXiv