MemCompile
A query-conditioned evidence compilation framework for long-term agent memory that decouples memory addressing from evidence presentation, compiling answer-ready context under budget constraints through an addressing plane and an evidence plane.
Timeline (last 90 days)
MemCompile is a framework that recovers answerable evidence from long-term memory through query-conditioned evidence compilation.
MemCompile's core design is divided into an Address Plane and an Evidence Plane, which are connected through a unified identity identifier.
A SourcePacket is the basic unit in MemCompile representing an evidence packet; it normalizes multiple heterogeneous sources from a single retrieval hit into a unified identity and retains source links for traceability.
MemCompile achieved 80.52% Judge Accuracy on the LoCoMo dataset.
MemCompile extends to higher quality ranges on the quality-cost Pareto frontier, meaning it can provide better answer quality at the same cost, or lower cost at the same quality.
MemCompile also shows strong performance on the LongMemEval-S benchmark, indicating that the benefits of query-conditioned compilation generalize to different long-term conversational memory tasks.
Decoupling memory addressing from evidence presentation provides a more controllable and efficient paradigm for long-term agent memory.
Performing lossy but controllable compilation at the evidence level can actually improve the accuracy of the final answer.
MemCompile's code and implementation are not yet open source; if open-sourced later, it could become a strong baseline for long-term memory systems.
All Facts (9)
MemCompile is a framework that recovers answerable evidence from long-term memory through query-conditioned evidence compilation.
50%UnverifiedMemCompile's core design is divided into an Address Plane and an Evidence Plane, which are connected through a unified identity identifier.
50%UnverifiedA SourcePacket is the basic unit in MemCompile representing an evidence packet; it normalizes multiple heterogeneous sources from a single retrieval hit into a unified identity and retains source links for traceability.
50%UnverifiedMemCompile achieved 80.52% Judge Accuracy on the LoCoMo dataset.
50%UnverifiedMemCompile extends to higher quality ranges on the quality-cost Pareto frontier, meaning it can provide better answer quality at the same cost, or lower cost at the same quality.
50%UnverifiedMemCompile also shows strong performance on the LongMemEval-S benchmark, indicating that the benefits of query-conditioned compilation generalize to different long-term conversational memory tasks.
50%UnverifiedDecoupling memory addressing from evidence presentation provides a more controllable and efficient paradigm for long-term agent memory.
50%UnverifiedPerforming lossy but controllable compilation at the evidence level can actually improve the accuracy of the final answer.
50%UnverifiedMemCompile's code and implementation are not yet open source; if open-sourced later, it could become a strong baseline for long-term memory systems.
50%