Researchers introduce APDMem hierarchical memory for long-context LLM assistants
A new paper on arXiv proposes APDMem, a hierarchical long‑term memory system designed for personalized LLM assistants that must pull sparse evidence from lengthy conversation histories. The architecture splits the dialogue into four layers – thematic summaries, personalized key facts, turn‑level evidence notes, and raw messages – allowing the system to treat memory as a progressive disclosure…
Key points
- APDMem organizes conversation history into four progressive layers: summaries, key facts, turn‑level notes, and raw messages
- A controller applies progressive disclosure, reading high‑level summaries first and drilling deeper only when needed
- On LongMemEval, APDMem accesses only 8% of conversation data yet achieves strong long‑context reasoning performance
During inference a controller first reads the high‑level summaries and only drills down to finer layers when the query demands more detail. This adaptive approach creates a cost‑fidelity trade‑off: simple questions can be answered after a brief scan, while complex, multi‑hop or exact‑evidence queries trigger deeper inspection of the lower layers. A note synthesizer then formats the retrieved evidence into a structured answer, flagging contradictions before the final response.
Experiments on the LongMemEval benchmark show that APDMem attains strong long‑context reasoning performance while accessing just 8% of the total conversation data, demonstrating that progressive disclosure can dramatically reduce compute without sacrificing accuracy.
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- Google VP Yossi Matias says AI’s biggest impact may come from intersecting fields · 1 src
- HakemBench releases 2,346-item Turkish benchmark for typed decisions · 1 src
- Researchers introduce MEA, a Reward-Driven Multi-Agent system for faithful model explanations · 1 src
- Tropical reinforcement learning algorithm tropic improves compositional reasoning · 1 src
- Claude Opus meta-agent achieves 81.3% mean pass@2 on generated terminal tasks · 1 src
Comments
via GitHub Discussions