AuthorsZhen Bi, Xueshu Chen, Yan Wang et al.
TL;DR
Memory Boundary-Aware Router uses behavioral knowledge-boundary signals to selectively route conditional memory, yielding 0.63 ERR accuracy on BioProBench vs 0.62 for Qwen3-8B-Memory-LoRA (+0.01).
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Conditional memory is supposed to help scientific reasoning, yet memory-enabled Qwen3-8B-Memory-LoRA drops Protocol QA accuracy to 0.39 and Brier score rises to 0.50.
On ChemCoTBench, enabling memory causes prediction-instability cohorts to regress by 4.45 and 3.81 percentage points, showing that naive memory activation can disrupt reasoning and calibration.
HOW IT WORKS
Memory Boundary-Aware Router builds on Behavioral Scientific Knowledge Boundary Characterization, Knowledge-Circuit View of Internal Knowledge Boundaries, External Boundary-Aware Data Router, and Internal Boundary-Aware Parameter Router to control memory.
Think of Memory Boundary-Aware Router as a smart memory controller: the external router decides when to access memory, while the internal router chooses where inside the Transformer to inject it, like selective RAM access into specific CPU stages.
This KEY_MECHANISM lets Memory Boundary-Aware Router activate conditional memory only for inputs and layer–stage nodes where it helps, avoiding the interference that a plain context window or always-on memory cannot prevent.
DIAGRAM
This diagram shows how Memory Boundary-Aware Router connects behavioral boundary codes with internal knowledge-circuit nodes during scientific reasoning.
DIAGRAM
This diagram shows how Memory Boundary-Aware Router is evaluated on BioProBench and ChemCoTBench, including random routing and layer ablations.
PROCESS
Behavioral Scientific Knowledge Boundary Characterization
Memory Boundary-Aware Router first characterizes answer failure, reasoning-step failure, and prediction instability using Behavioral Scientific Knowledge Boundary Characterization to understand where the backbone struggles.
Knowledge-Circuit View of Internal Knowledge Boundaries
Memory Boundary-Aware Router then maps hidden states to layer–stage knowledge-circuit nodes via the Knowledge-Circuit View of Internal Knowledge Boundaries, defining where memory can intervene.
External Boundary-Aware Data Router
Using task-specific input proxies, the External Boundary-Aware Data Router computes ρt(x) and gate πt(x) to decide whether conditional memory should be activated for the query.
Internal Boundary-Aware Parameter Router
Finally, the Internal Boundary-Aware Parameter Router selects memory-enabled layers, stages, and node strengths ai,l,s so Memory Boundary-Aware Router injects memory only at helpful knowledge-circuit nodes.
KEY CONTRIBUTIONS
Systematic empirical study of conditional memory in scientific reasoning
Memory Boundary-Aware Router is grounded in a study across BioProBench and ChemCoTBench, two backbone families, and six task types, showing memory repairs some failures but induces regressions in others.
Behavioral and knowledge-circuit characterization of memory utility
Memory Boundary-Aware Router uses Behavioral Scientific Knowledge Boundary Characterization and the Knowledge-Circuit View of Internal Knowledge Boundaries to show utility depends jointly on input, injection location, and contribution strength.
Knowledge Boundary-Aware Router for selective memory activation
Memory Boundary-Aware Router introduces External Boundary-Aware Data Router and Internal Boundary-Aware Parameter Router that beat activation-rate-matched random routing and avoid regressions from static memory configurations.
RESULTS
Accuracy
0.63
+0.01 over Qwen3-8B-Memory-LoRA on BioProBench Error Correction
Accuracy
0.60
+0.03 over Qwen2.5-7B-Memory-LoRA on BioProBench Error Correction
Accuracy
0.51
+0.05 over Qwen2.5-7B-Memory on ChemCoTBench MolEdit
MAE
0.50
-0.11 vs Qwen2.5-7B-Memory on ChemCoTBench MolUnd (lower is better)
On BioProBench, which tests biological procedural reasoning (Error Correction, Step Ordering, Protocol Question Answering), Memory Boundary-Aware Router improves ERR accuracy to 0.63 with Qwen3-8B and 0.60 with Qwen2.5-7B. On ChemCoTBench, which evaluates molecular editing, optimization, and understanding, Memory Boundary-Aware Router raises MolEdit accuracy to 0.51 and reduces MolUnd MAE to 0.50, demonstrating that selective routing preserves beneficial memory while suppressing harmful effects.
BENCHMARK
On BioProBench, which tests biological procedural reasoning (Error Correction, Step Ordering, Protocol Question Answering), Memory Boundary-Aware Router improves ERR accuracy to 0.63 with Qwen3-8B and 0.60 with Qwen2.5-7B. On ChemCoTBench, which evaluates molecular editing, optimization, and understanding, Memory Boundary-Aware Router raises MolEdit accuracy to 0.51 and reduces MolUnd MAE to 0.50, demonstrating that selective routing preserves beneficial memory while suppressing harmful effects.
BENCHMARK
Accuracy on BioProBench Error Correction for Qwen3-8B variants and Memory Boundary-Aware Router.
BENCHMARK
Editing accuracy on ChemCoTBench MolEdit for Qwen2.5-7B variants and Memory Boundary-Aware Router.
KEY INSIGHT
Enabling memory on ChemCoTBench prediction-instability cohorts causes performance drops of 4.45 and 3.81 percentage points instead of improvements.
This is counterintuitive because conditional memory is expected to help unstable inputs, yet Memory Boundary-Aware Router shows that naive activation can worsen scientific reasoning.
WHY IT MATTERS
Memory Boundary-Aware Router unlocks conditional memory as a selective computational resource, tuned per task, backbone, and input regime rather than always on.
Builders can now design scientific reasoning systems that route memory only when behavioral and knowledge-circuit signals predict benefit, avoiding regressions that previously made memory augmentation unreliable.
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.
Shannan Yan, Jingchen Ni et al.
· 2026
AdaMem organizes dialogue history into Working Memory, Episodic Memory, Persona Memory, and Graph Memory coordinated by a Memory Agent, Research Agent, and Working Agent. On LoCoMo with GPT-4.1-mini, AdaMem achieves 44.65 F1 overall, beating the best baseline LangMem at 41.76 F1 by +2.89.
Answers use this explainer on Memory Papers.
Checking…