conversational-long-term-memory-8-14b-backbones-plain-top-k-dense-retrieval CLM-0371For conversational long-term memory with 8-14B backbones and plain top-K dense retrieval, writing each dialogue span into several complementary low-overlap semantic views before indexing retrieves gold evidence for more queries than structured extraction into a single view, and the advantage is largest at small retrieval budgets.
Capability: Remembering across sessions
Observed on
Qwen3-8B/14B on LoCoMo and PersonaMem-32k, shared e5-base-v2 embeddings, fixed top-30 per-store budget; evidence-level dedup by provenance.
Sources
- Validated by: LoCoMo overall LLM-Judge 0.837 vs 0.774 for single-view structured extraction and 0.683-0.783 for Mem0/MemGAS/MemoryBank/A-mem; no-positive rate 5.14% vs 5.53% single-view and 9.83% without consolidation; PersonaMem-32k accuracy 69.10 at 14B vs 63.33 best baseline. Measured on two benchmarks (1,540 LoCoMo questions, 589 PersonaMem items) with two backbones and four memory baselines; ablations separate multi-view organization, DPP selection and consolidation. Judge is Qwen3-8B, same family as generators, so judge bias is not isolated; [truncated]
Disagreeing is the most useful thing you can do here. Nothing on this page waits for a reviewer: the counts above move when evidence lands, and every source stays attached with who brought it.
Related claims
- For multi-hop QA over 28k-896k token haystacks with frozen Qwen2.5 backbones (7B-72B), summarizing the document in a parallel hierarchical tree — overlapping windows scanned ~5x, raw segments scored 'critical' passed straight to the final answer, and one summarization channel per decomposed sub-question — scores higher than sequential memory-agent and BM25 retrieval baselines without training, and beats an RL-trained memory agent on the out-of-distribution dataset while losing to it in-distribution.Losing information in long inputs · reviewed by AI
- When accurate retrieved context is supplied at every turn of a three-turn dialogue, model scores still fall as each turn adds a condition or constraint on how the retrieved material must be used, and models that score alike on plain multi-turn chat diverge on retrieval-grounded turns, so using retrieval well is a distinct competence from chatting well.Using what was retrieved, and only that · reviewed by AI
- For spoken language models answering questions over long dialogue histories, converting the history to transcripts beats keeping it as audio when the answer is a lexical detail (exact slot values across dense turns: mean 71.3% vs 42.6%), but loses when the answer depends on how something was said, where audio history wins for six of seven models and by a larger margin once lexical emotion cues are removed.Understanding speech and audio · reviewed by AI
- Recursively summarizing the dialogue into a running memory improves consistency in long conversations.Remembering across sessions
- On long-context tasks over 8K-128K-token documents in the six official UN languages, models answer word-level span questions well (about 73% average accuracy across 12 models) but degrade sharply as the required unit coarsens to sentence selection (about 21%), paragraph filling and document-level summarization and translation, so long-context competence measured by needle-style word retrieval overstates comprehension of larger discourse units.Losing information in long inputs · reviewed by AI
Notes
Drafted from the paper by a model and filed unreviewed. Visible here so it can be read, not because anyone has vouched for it: it does not move any technique's standing and does not count toward the internal scorecard. Drafted confidence: low. Falsifier as drafted: A replication where single-view structured extraction with the same embeddings and budget matches or beats multi-view organization on no-positive rate and judged answer quality, or where the multi-view gain vanishes at small K.