Descriptive recall
A question resembles the memory.
“When did Caroline join the support group?” points to a fact that names the same group.
Entity triggers · Scene triggers
EMNLP 2026 MAIN CONFERENCE
T-MemMemory That Anticipates,
Not Archives.
Rehearsing tomorrow’s questions at write time.
Long-term memory for both descriptive and associative recall.
Accepted to EMNLP 2026 Main Conference · Camera-ready version submitted
80.26%
LoCoMo accuracy
74.81%
LoCoMo-Plus accuracy
5.45 pp
Cross-benchmark gap
Official benchmark protocols · LLM-as-judge accuracy · LoCoMo-Plus Cognitive subset
01 / The idea
Remember a fact.
Anticipate when it matters.
Similarity-based retrieval works well when a question resembles a stored memory. But a future question may be connected to a past conversation through a situation, a consequence, or a personal arc—without sharing its wording.
T-Mem prepares for these connections when memories are written. It creates triggers: descriptions of the cues and future situations under which a memory will be useful. Across two evidence granularities—atomic items and conversational scenes—four trigger families support both ways of remembering.
Associative recall
A situation makes the memory relevant.
A question about a team-dinner venue can make a colleague’s allergy, mentioned months earlier, matter again.
Bridge triggers · Horizon triggers
02 / The method
Construct once.
Retrieve from multiple cues.
T-Mem builds a typed graph of scenes, items, and topic labels. At read time, a topic → scene → item cascade combines lexical and dense rankings with reciprocal rank fusion. Trigger recall can bypass the topic prefilter, bringing relevant evidence into the candidate pool even when the query falls outside its similarity neighborhood.
Scenes & topics
Segment by event closure, then assign growing, multi-label topics.
Atomic & connected facts
Extract individual facts and preserve connections across scenes.
Entity & Bridge
Give each item descriptive concepts and future-use situations.
Scene & Horizon
Describe an exchange and project its relevance into future contexts.
These are conceptual construction stages. The reproduction scripts include indexing, trigger building (stage 4b), retrieval, persona extraction, and evaluation.
03 / Results
Two benchmarks.
Two ways of remembering.
The main comparison follows each benchmark’s official protocol. Memory construction uses GPT-4.1-mini and the dense encoder is bge-m3. LoCoMo uses GPT-4o-mini for answering and judging; LoCoMo-Plus uses GPT-4o for answering and Gemini-2.5-Flash as judge.
LoCoMo
LLM-judge accuracy (%) · token-level F1| Method | Single-hop | Multi-hop | Temporal | Open Domain | Overall | F1 |
|---|---|---|---|---|---|---|
| MIRIX | 68.22 | 54.26 | 68.54 | 46.88 | 64.33 | 28.10 |
| Mem0 | 73.33 | 58.75 | 52.34 | 45.83 | 64.57 | 43.46 |
| Zep | 66.23 | 52.12 | 54.82 | 33.33 | 59.22 | 41.23 |
| Memobase | 73.12 | 64.65 | 81.20 | 53.12 | 72.01 | 50.18 |
| MemU | 66.34 | 63.12 | 27.10 | 50.00 | 56.55 | 35.15 |
| Supermemory | 67.30 | 51.12 | 31.77 | 42.67 | 55.34 | 34.87 |
| MemOS | 81.09 | 67.49 | 75.18 | 55.90 | 75.80 | 45.27 |
| HyperMem | 84.34 | 62.29 | 79.44 | 47.92 | 77.01 | 49.93 |
| T-Mem | 85.97 | 69.15 | 82.55 | 55.21 | 80.26 | 51.96 |
Main-paper protocol, averaged over three runs. T-Mem improves Overall by 3.25 percentage points over HyperMem. Baseline numbers other than HyperMem are from Li et al. (MemOS, 2025), as reported in the paper.
LoCoMo → LoCoMo-Plus
Associative recall · lower gap is better| Method | LoCoMo (%) | LoCoMo-Plus (%) | Gap (pp) ↓ |
|---|---|---|---|
| GPT-4o full context | 66.14 | 21.05 | 45.09 |
| Gemini-2.5-Pro full context | 71.46 | 26.06 | 45.40 |
| Mem0 | 64.94 | 15.80 | 49.14 |
| SeCom | 64.97 | 14.90 | 50.07 |
| A-Mem | 66.70 | 17.20 | 49.50 |
| MemOS | 75.80 | 32.67 | 43.13 |
| HyperMem | 77.01 | 48.63 | 28.38 |
| T-Mem | 80.26 | 74.81 | 5.45 |
LoCoMo-Plus denotes the Cognitive subset. The full-context baselines use the complete raw dialogue without a memory mechanism. Mem0, SeCom, A-Mem and full-context numbers are from Li et al. (LoCoMo-Plus, 2026); this table follows the paper’s cross-benchmark comparison.
Additional camera-ready experiments Model sensitivity & QA protocol
Stronger build and answer models
With the judge fixed at GPT-4o-mini, GPT-5.1 for both memory construction and answering reaches 84.85% LoCoMo Overall accuracy. These configurations vary the build and QA models and are reported separately from the main benchmark above.
| Build model | QA model | Overall (%) |
|---|---|---|
| GPT-4.1 | GPT-4o-mini | 78.31 |
| GPT-4.1 | GPT-4.1 | 82.51 |
| GPT-5.1 | GPT-4o-mini | 81.36 |
| GPT-5.1 | GPT-5.1 | 84.85 |
Using open-weight Qwen3-32B for construction, with QA and judge fixed at GPT-4o-mini, yields 75.45% (versus 80.26% for GPT-4.1-mini construction).
Matching HyperMem’s QA protocol
Under HyperMem’s GPT-4.1-mini + seven-step CoT answer-generation protocol, T-Mem scores 93.70% accuracy / 16.97 F1, compared with HyperMem’s reported 92.73% / 15.78. The prompt produces much longer answers, so these scores are not directly comparable with the official-protocol table.
04 / Analysis
What makes the memory work?
05 / Explore the pipeline
Follow a conversation
into memory.
Inspect stored outputs from a real LoCoMo conversation: raw turns, scenes, extracted facts, all four trigger families, and retrieved answers. The explorer uses pre-computed artifacts; the Run locally view explains how to reproduce the pipeline.
06 / Citation
Build on this work.
Accepted to the EMNLP 2026 Main Conference. The arXiv citation below remains available while the proceedings record is being published.
BibTeX
@misc{guo2026tmem,
title = {T-Mem: Memory That Anticipates, Not Archives},
author = {Guo, Weidong and Wang, Dakai and Wang, Zixuan and Liu, Hui and Xu, Yu},
year = {2026},
eprint = {2606.15405},
archivePrefix = {arXiv},
primaryClass = {cs.CL},
note = {Accepted to EMNLP 2026 Main Conference},
url = {https://arxiv.org/abs/2606.15405}
}