Group-chat memory stops ranking by similarity alone: find the burst of discussion, then link updates and support into an evidence forest

Memory retrieval for long multi-party chats no longer ranks only by semantic similarity; temporal structure is used both to recover evidence and to reconstruct how an event evolved. Earlier temporally aware memory frameworks mostly store time as metadata or use temporal signals to guide retrieval over preconstructed memory structures, without using temporal structure to both recover locally concentrated evidence and reconstruct the complete event progression. On GroupMemBench, T-Mem reaches 63.09% and 62.68% overall accuracy with GPT-5.5 and Deepseek-V4-Flash, 8.59 and 8.05 points above the strongest baseline; on EverMemBench it reaches 79.46% and 77.63%, 8.42 and 6.50 points above the strongest completed baseline.

Retrieval first finds a few highly relevant messages as temporal anchors, then re-examines the discussion episodes where those anchors concentrate to select evidence again. Global similarity ranking misses the update that decides the current state when it shares few words with the query, while simply raising Top-k pulls in unrelated episodes and competing topics. On GroupMemBench at Top-36, key-evidence recall rises from 56.6% for BM25 and 61.5% for HippoRAG2 to 72.8% for T-Mem; even at Top-48, HippoRAG2 reaches only 62.5% against 75.6% for T-Mem.

Recovered messages are organized into a query-conditioned evidence forest ordered in time, with Update edges for state progression and Expand edges for supporting elaboration. Given the same messages as a flat chronological list, the answer model must work out which message is the current state and which has been superseded. In a comparison holding the retrieved messages identical and changing only the organization, accuracy improves at every retrieval depth, with gains rising from 3.09 points at Top-24 to 7.60 points at Top-48; the flat list peaks at 50.74% at Top-36 and then degrades, while the organized condition peaks later at 55.70% with Top-48.

Evaluation covers two long-term multi-party memory benchmarks, GroupMemBench and EverMemBench, with GPT-5.5 and Deepseek-V4-Flash as answer and judge backbones and answer accuracy as the metric. The evidence-recovery experiment on GroupMemBench excludes abstention questions, since the absence of answer-bearing evidence is their defining property. The metadata ablation runs under Deepseek-V4-Flash, where removing topic and phase fields leaves 60.40% overall accuracy.

Temporal evidence discovery relies on initially retrieved messages as anchors; with a small initial budget the anchor set may be too sparse or miss the relevant episode, limiting what temporal expansion can recover. The evidence forest approximates event evolution with only Update and Expand edges, so questioning, clarification, justification and elaboration may all fall under Expand. Both benchmarks provide cleaner metadata than many deployed group-chat systems, so generalization to noisier real conversations needs further study; answer generation and judging are both LLM-based and may vary across repeated runs.