AuthorsYu Cheng, Jiuan Zhou, Yongkang Hu et al.
TL;DR
TAME couples an Executor–Evaluator loop over a trust-aware memory bank, boosting GPT-5.2 AIME accuracy to 73.3% (+14.6pp over ReasoningBank) while maintaining competitive trustworthiness.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Existing memory evolution keeps only strategies with high task reward, so capability grows while trustworthiness drops, violating E[rtrust] ≥ τ.
Under benign task streams, agents accumulate shortcut strategies Stoxic, and Eq. (3) shows lim t→∞ P(s ∈ Stoxic | M(t)) → 1, causing deployment time reward hacking and trust drift.
HOW IT WORKS
TAME’s core mechanism is a Strategy Memory Bank M(t) with trust aware memory items m(t) i, an Executor, an Evaluator, and feedback driven memory evolution.
Think of TAME like RAM plus a safety aware cache: the Executor reads high usefulness strategies, while the Evaluator writes back calibrated usefulness and trust annotations, avoiding toxic shortcuts.
This KEY_MECHANISM lets TAME selectively reinforce, cautiously reuse, and continuously expand memory, achieving trustworthy test time evolution beyond what a plain context window or score only memory can do.
DIAGRAM
This diagram shows how TAME’s Executor and Evaluator interact over one query to retrieve, use, and update trust aware memories.
DIAGRAM
This diagram shows how Trust-Memevo streams evolution data and then freezes memory to evaluate TAME’s task utility and multi dimensional trustworthiness.
PROCESS
Trust aware memory bank construction
TAME represents each memory item as m(t) i = (qi, ei, Q(t) i, ψi), initializing Qinit = 0.5 and attaching trust annotations ψi to extracted experiences.
Executor retrieval and response generation
Given q(t), the Executor builds C(t) by sim(q(t), qi) ≥ δsim, selects R(t) via TopK on Q(t) i, and generates a(t) conditioned on ei and ψi.
Evaluator contribution scoring and trust annotation
The Evaluator assigns contribution scores c(t) i ∈ [0,1] to each m(t) i ∈ R(t) and produces a new trust annotation ψ(t) for the trajectory experience e(t).
Feedback driven memory evolution
TAME updates usefulness with Q(t+1) i = Q(t) i + α c(t) i (rtask − Q(t) i) and appends (q(t), e(t), Qinit, ψ(t)) to M(t+1), calibrating future reuse.
KEY CONTRIBUTIONS
Trust-Memevo benchmark construction
TAME is evaluated on Trust-Memevo, which pairs evolution sets with trustworthiness sets of 946 science, 700 math, 385 code, and 298 tool use instances.
Trust aware memory evolution framework
TAME introduces a shared Strategy Memory Bank M(t) with trust annotated memory items and an Executor–Evaluator loop for selective reinforcement and cautious reuse.
Empirical validation across domains
TAME shows strong utility and trust on MMLU-Pro, GPQA, GSM8K, MATH, AIME, TaskBench, and LiveCodeBench, including +0.146 AIME accuracy over ReasoningBank on GPT-5.2.
RESULTS
AIME accuracy
0.733
+0.146 over ReasoningBank on GPT-5.2
GSM8K accuracy
0.968
0.019 over ReasoningBank on GPT-5.2
MATH accuracy
0.909
0.030 over ReasoningBank on GPT-5.2
MMLU-Pro accuracy
0.888
-0.003 vs ReasoningBank on GPT-5.2 while improving trust balance
On Trust-Memevo’s math and science tracks, including AIME, GSM8K, MATH, and MMLU-Pro, these numbers show that TAME can raise task accuracy substantially while keeping trustworthiness competitive compared to ReasoningBank and safety guarded variants.
BENCHMARK
On Trust-Memevo’s math and science tracks, including AIME, GSM8K, MATH, and MMLU-Pro, these numbers show that TAME can raise task accuracy substantially while keeping trustworthiness competitive compared to ReasoningBank and safety guarded variants.
BENCHMARK
Accuracy on historical AIME problems under test time memory evolution.
BENCHMARK
Accuracy on the MATH dataset with different memory evolution methods.
KEY INSIGHT
TAME raises GPT-5.2 AIME accuracy from 0.587 with ReasoningBank to 0.733, yet maintains competitive trustworthiness scores instead of degrading them.
This is surprising because prior utility driven memory evolution typically improves accuracy while lowering trust compliance, but TAME’s trust aware updates break that assumed trade off.
WHY IT MATTERS
TAME unlocks agents that can evolve their memory at test time, gaining complex reasoning skills while explicitly tracking usefulness and trust for each experience.
Builders can now deploy continually learning agents on math, science, code, and tool use streams without accepting unchecked safety drift from shortcut strategies accumulating in memory.
Peilin Feng, Suorong Yang, Soujanya Poria
arXiv 2026 · 2026
Σ-Mem maintains historical competence evidence matrices M_p, a peer relationship matrix G, a memory readout r_{p,t}, and residual steering δ_{p,t} to guide multi-agent trust decisions. On the mixed counterfactual benchmark, Σ-Mem raises Qwen3-0.6B accuracy at CF@90 from 46.22% with the base model to 71.10% with Σ w/ G.
Nikhil Verma
· 2026
Focus Agent adds start_focus, complete_focus, a persistent Knowledge block, and an optimized Persistent Bash plus String-Replace Editor scaffold to actively compress context during long software-engineering tasks. On five hard SWE-bench Lite instances against a Baseline ReAct agent, Focus Agent achieves 22.7% token reduction (14.9M → 11.5M) while matching 3/5 = 60% task success.
Xiaohui Zhang, Zequn Sun et al.
· 2026
ActMem transforms dialogue history into atomic facts via Memory Fact Extraction, groups them with Fact Clustering, links them through a Memory KG Construction module, and uses Counterfactual-based Retrieval and Reasoning for action-aware answers. On ActMemEval, ActMem reaches 76.52% QA accuracy with DeepSeek-V3, beating LightMem’s 63.97% by 12.55 points and NaiveRAG’s 61.54%.
Answers use this explainer on Memory Papers.
Checking…