AuthorsBoyu Yang, Jiazheng Sun, Zilong Lu et al.
TL;DR
MeClear uses cooperative Shapley-style memory attribution plus verified query-scoped clearance to reach 82.3% task recovery and 85.9% target recall on long-dialogue memory faults.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Long-horizon agents using semantic retrieval often pull stale, misleading, or conflicting memories, causing failures that Leave-One-Out only recovers 56.8% of the time.
When these faulty memories interact redundantly or jointly, long-dialogue LLM agents misanswer user queries, degrading downstream task utility and propagating errors across future tasks.
HOW IT WORKS
MeClear combines Adaptive Task-Conditioned Context Retrieval, Local Counterfactual Screening, Counterfactual Cooperative Attribution, and Verified Query-Scoped Clearance to manage harmful memories without editing the database.
You can think of MeClear like a CPU with RAM and a debugger: retrieval loads candidate memories, Shapley-style attribution debugs which lines are harmful, and clearance sets breakpoints that hide only the bad parts.
This cooperative attribution lets MeClear uncover redundant and joint harm that a plain context window or single-record deletion cannot, enabling minimal, reversible context clearance per query.
DIAGRAM
This diagram shows how MeClear processes a single query from retrieval through Leave-One-Out screening, Shapley attribution, and nested clearance verification.
DIAGRAM
This diagram shows how MeClear is evaluated on LoCoMo conversations across direct conflicts, redundant conflicts, and joint interactions using Kimi k2.6 and Qwen3.6 Flash.
PROCESS
Adaptive Task-Conditioned Context Retrieval
MeClear uses Adaptive Task-Conditioned Context Retrieval to freeze a query-specific context Mt from the external memory bank before any attribution.
Local Counterfactual Screening
MeClear runs Local Counterfactual Screening by computing Leave-One-Out effects d_t,i on Mt to get an interpretable local contribution profile.
Counterfactual Cooperative Attribution
MeClear performs Counterfactual Cooperative Attribution with permutation sampled Shapley estimates ψ_t,i and optional pairwise interaction profiles ω_t,ij.
Verified Query-Scoped Clearance
MeClear executes Verified Query-Scoped Clearance by testing nested clearance candidates C_t,j and selecting the minimal-cardinality context that maximizes gain b_g_t(j).
KEY CONTRIBUTIONS
Task Conditioned Memory Clearance Framework
MeClear introduces a task conditioned memory clearance framework that unifies Adaptive Task-Conditioned Context Retrieval, Counterfactual Cooperative Attribution, and Verified Query-Scoped Clearance for long-horizon agents.
Cooperative Memory Attribution Methodology
MeClear develops a cooperative memory attribution methodology that integrates Local Counterfactual Screening with permutation sampled Shapley estimation to resolve redundant conflict masking and multi memory interactions.
Comprehensive Long Dialogue Evaluation
MeClear is evaluated on ten long dialogue memory pools with 745 test cases and 1,115 fault records, achieving 85.9% target recall and 82.3% task recovery, a +25.5 percentage point gain over Leave-One-Out.
RESULTS
Recall@|M| memory micro
85.9%
+47.6 percentage points over Leave-One-Out
Binary recovery case macro
82.3%
+25.5 percentage points over Leave-One-Out
Exact H = M case macro
47.0%
+12.8 percentage points over Leave-One-Out
Complete M ⊆ H case macro
86.7%
+22.5 percentage points over Leave-One-Out
These metrics come from 745 causally verified cases over ten LoCoMo long dialogue conversations, testing memory fault detection and task recovery. The gains show that MeClear reliably finds harmful coalitions and restores execution where Leave-One-Out and other baselines fail under redundant and joint interactions.
BENCHMARK
These metrics come from 745 causally verified cases over ten LoCoMo long dialogue conversations, testing memory fault detection and task recovery. The gains show that MeClear reliably finds harmful coalitions and restores execution where Leave-One-Out and other baselines fail under redundant and joint interactions.
BENCHMARK
Binary recovery case macro on long dialogue memory pools.
KEY INSIGHT
Under redundant conflicts, Leave-One-Out recall collapses to 12.3% and recovery to 6.3%, while MeClear maintains 83.2% recall and 68.4% recovery.
This is surprising because single-record deletion is often assumed sufficient, but MeClear shows that coalition-aware Shapley attribution is necessary to expose masked multi-memory harm.
WHY IT MATTERS
MeClear unlocks query-scoped, reversible memory clearance that respects cooperative interactions, preventing harmful records from derailing long-horizon agents without editing persistent storage.
Builders can now deploy agents that accumulate large experience pools while dynamically suppressing only the minimal harmful coalitions per query, even under redundant and joint conflicts.
Jingbiao Mei, Jinghong Chen et al.
arXiv 2026 · 2026
ATM-Bench structurally evaluates long-term multimodal personal memory using Memory Ingestion, Retrieval, and Answer Generation with Schema-Guided Memory and Descriptive Memory variants. On ATM-Bench-Hard, Oracle with SGM reaches 47.3% QS while the best full system stays under 20% accuracy, revealing a large gap.
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.
Xingyu Lyu, Jianfeng He et al.
· 2026
ADAM combines Anchor extraction, Distribution estimation, Anchor selection, and Query generation to adaptively probe agent memory via an auxiliary generator and entropy based selection. On the EHRAgent benchmark with Llama2-7b-chat, ADAM reaches EQ=77 and ASR=1.00, compared to MEXTRA’s EQ=44 and ASR=0.89.
Answers use this explainer on Memory Papers.
Checking…