APDMem
An arXiv research paper on AI memory architecture that proposes a four-layer progressive memory structure with an adaptive retrieval mechanism. A controller selectively traverses memory layers on demand to enable accurate and efficient retrieval of long-term conversational memory.
Timeline (last 90 days)
APDMem (Agent-controlled Progressive Disclosure Memory) is a memory architecture for long-term memory retrieval that employs a hierarchical retrieval mechanism.
APDMem organizes conversation history into four progressively refined levels: topic summaries, personalized key facts, turn-level evidence notes, and raw messages.
During inference, APDMem uses a controller that first reads high-level summaries and drills down to finer-grained evidence only when needed, achieving progressive disclosure.
APDMem introduces a note synthesizer that converts scattered retrieval results into query-centric structured representations, including integrating relevant facts, ordering events, and flagging contradictions between pieces of evidence.
APDMem was tested on the LongMemEval benchmark, achieving strong performance on long-context memory reasoning tasks while accessing only 8% of the full conversation.
APDMem achieves a cost-fidelity trade-off through adaptive retrieval: simple queries can terminate early at a shallow level, while complex queries trigger deeper-level inspection.
APDMem was published on arXiv as a newly released preprint.
Traditional memory solutions rely on flat storage or fixed retrieval granularity, resulting in either too coarse a granularity to find sufficient detail or too fine a granularity that wastes computational cost.
All Facts (8)
Traditional memory solutions rely on flat storage or fixed retrieval granularity, resulting in either too coarse a granularity to find sufficient detail or too fine a granularity that wastes computational cost.
50%UnverifiedAPDMem organizes conversation history into four progressively refined levels: topic summaries, personalized key facts, turn-level evidence notes, and raw messages.
50%UnverifiedDuring inference, APDMem uses a controller that first reads high-level summaries and drills down to finer-grained evidence only when needed, achieving progressive disclosure.
50%UnverifiedAPDMem introduces a note synthesizer that converts scattered retrieval results into query-centric structured representations, including integrating relevant facts, ordering events, and flagging contradictions between pieces of evidence.
50%UnverifiedAPDMem was tested on the LongMemEval benchmark, achieving strong performance on long-context memory reasoning tasks while accessing only 8% of the full conversation.
50%UnverifiedAPDMem achieves a cost-fidelity trade-off through adaptive retrieval: simple queries can terminate early at a shallow level, while complex queries trigger deeper-level inspection.
50%UnverifiedAPDMem (Agent-controlled Progressive Disclosure Memory) is a memory architecture for long-term memory retrieval that employs a hierarchical retrieval mechanism.
50%UnverifiedAPDMem was published on arXiv as a newly released preprint.
50%