AuthorsDong Xu, Zhangfan Yang, Jiantao Wu et al.
TL;DR
RoutePrism uses paired construction paths and a four-condition support restoration test to show that restoring a single displaced record recovers over 60 percentage points of lost accuracy.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Processing the same records in a different order can discard different evidence, yet endpoint accuracy alone cannot reveal what changed or whether it mattered.
Memory policies in long-horizon assistants silently drop the only supporting record for a query, causing incorrect answers while standard benchmarks report a single conflated accuracy score.
HOW IT WORKS
RoutePrism combines Task-preserving eligibility, Paired Memory Construction, Observable Diagnostics, and Support Intervention and Analysis to isolate construction-order effects under fixed policies and answer models.
Think of RoutePrism as running two different ingestion orders into the same "memory card catalog", then swapping one card back in to see if the catalog still answers correctly.
This paired-path plus four-condition design lets RoutePrism pinpoint when survivor selection or summarization loses task-relevant evidence that a plain context window view of accuracy cannot expose.
DIAGRAM
This diagram shows how RoutePrism runs eligibility checks, builds paired memories, computes diagnostics, and performs the four-condition support restoration intervention.
DIAGRAM
This diagram shows how RoutePrism evaluates different memory policies on PersonaMem-32K and LongMemEval-S with multiple answer models.
PROCESS
Stage 1 Task-preserving eligibility
RoutePrism screens each question using Task-preserving eligibility to ensure permutations of the record pool preserve the same correct answer before any construction.
Stage 2 Paired Memory Construction
RoutePrism runs Paired Memory Construction, feeding the same source pool in forward and reversed orders to the memory policy while recording exposed sources and compiled contexts.
Stage 3 Observable Diagnostics
RoutePrism computes Observable Diagnostics, including J source overlap, H answer disagreement, and Δ correctness change, plus residual route effects using repeated answer draws.
Stage 4 Support Intervention and Analysis
RoutePrism performs Support Intervention and Analysis with four conditions S+, S−, Res., and PL to test whether restoring a displaced supporting record recovers task accuracy.
KEY CONTRIBUTIONS
Diagnostic protocol for construction-order effects
RoutePrism introduces a paired construction protocol with Task-preserving eligibility and Paired Memory Construction, showing Dres > 0.4 residual route effects beyond answer-model sampling noise.
Evidence that displaced records carry task value
RoutePrism’s Support Intervention and Analysis shows that restoring a single displaced record recovers over 60 percentage points of lost accuracy, while equal-length non-support replacements do not.
Distinct failure signatures across memory policies
RoutePrism reveals that compaction, bounded recency, MemoChat-style summarization, and A-MEM each exhibit different source, context, and metadata failure signatures that no single observation layer can capture.
RESULTS
Residual Dres focal
0.529
+0.036 over Recent-3 Dres 0.493
Correctness shift focal
0.444
+0.416 over Recent-3 correctness shift 0.028
Support R minus A focal
0.671
Res. 0.881 vs Abs. 0.210 on PersonaMem primary
Support R minus P Recent-3
0.750
Res. 0.917 vs Repl. 0.167 on PersonaMem primary
On PersonaMem-32K fact-recall queries, RoutePrism shows that reversing construction order under the focal compactor changes exposed sources (J = 0.057) and yields Dres = 0.529 residual route disagreement. The four-condition support restoration on PersonaMem demonstrates that RoutePrism can attribute over 60 percentage points of accuracy loss directly to displaced supporting records rather than prompt length or source count.
BENCHMARK
On PersonaMem-32K fact-recall queries, RoutePrism shows that reversing construction order under the focal compactor changes exposed sources (J = 0.057) and yields Dres = 0.529 residual route disagreement. The four-condition support restoration on PersonaMem demonstrates that RoutePrism can attribute over 60 percentage points of accuracy loss directly to displaced supporting records rather than prompt length or source count.
BENCHMARK
Answer accuracy in support present, absent, restored, and replacement conditions under focal and Recent-3 policies.
KEY INSIGHT
RoutePrism shows that restoring a single displaced record can recover over 60 percentage points of lost accuracy under both compaction and bounded recency.
This is surprising because many memory systems assume that as long as some related records remain, construction order and survivor selection should not drastically change task performance.
WHY IT MATTERS
RoutePrism gives builders a way to audit memory policies at the evidence level, separating source selection, context compilation, and answer sampling effects.
With RoutePrism, developers can design and regression-test survivor rules, summarizers, and metadata updates to avoid silent evidence loss that was previously invisible behind a single accuracy number.
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.
Xiaoyang Li, Yiqi Wang et al.
arXiv 2026 · 2026
Correlated Promotion Benchmark (CPB) combines CPB-Static, CPB-Live, a gold admission rule, lineage collapse, and a governance rule to stress-test epistemic admission in shared agent memory. On CPB-Live, the governance rule keeps damage shares between 0.112 and 0.152 and false adoption between 0.06 and 0.09, while majority vote and LLM judges often match share-all’s false adoption.
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.
Answers use this explainer on Memory Papers.
Checking…