AuthorsWenzhi Li, Dong Nie, Rui Lan et al.
TL;DR
Dual-Layer Agentic Memory uses a small-to-large write router plus slow parametric consolidation to prune 68% redundant memory while retaining 98.3% QA EM of Full Store.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Existing agent memories treat external memory as a monotonically growing repository, causing memory bloat and degraded retrieval precision over time.
In long-horizon LLM agents, this leads to noisy retrieval, higher latency, and rising compute costs, even when many stored facts are redundant or already in parametric memory.
HOW IT WORKS
Dual-Layer Agentic Memory introduces an operational memory taxonomy, fast write router, small-to-large cost-aware routing cascade, and write-back consolidation to manage what gets stored and internalized.
Conceptually, Dual-Layer Agentic Memory mirrors a hippocampus–neocortex system: external memory E acts as a fast editable buffer, while parametric memory Θ is a slower, stable substrate.
This dual-layer design lets Dual-Layer Agentic Memory selectively externalize high-value facts and later internalize them via supervised fine-tuning, avoiding permanent dumps that a plain context window cannot manage.
DIAGRAM
This diagram shows how Dual-Layer Agentic Memory routes each incoming fact through the small-to-large cascade and decides whether to write or discard.
DIAGRAM
This diagram shows how Dual-Layer Agentic Memory is evaluated on the online streaming ZsRE benchmark with write-time admission and downstream QA.
PROCESS
Operational Memory Taxonomy
Dual-Layer Agentic Memory first labels each fact as non-write, write-new, or write-update by comparing zero-shot and memory-supported answers, defining admission targets for the fast write router.
Small-to-Large Cost-Aware Write Routing
Dual-Layer Agentic Memory uses the small-to-large routing cascade with qwen3-1.7B and qwen3-8B plus an escalation gate trained on reward differences to decide write versus discard.
Online Streaming Knowledge Lifecycle Benchmark
Dual-Layer Agentic Memory processes 300 episodes with 500 turns and 100 fact injections per episode, making write-time decisions before queries and then answering downstream QA with or without external memory.
Write-Back as Slow Consolidation
Dual-Layer Agentic Memory periodically converts write-new and write-update facts into supervised fine-tuning examples, validates internalization with factual probes, and evicts consolidated facts from external memory.
KEY CONTRIBUTIONS
Dual-Layer Knowledge Lifecycle System
Dual-Layer Agentic Memory unifies parametric memory Θ, external memory E, fast write routing, and slow write-back consolidation into a single lifecycle, evaluated on 300 streaming episodes with 500 turns each.
Operational Memory Taxonomy and Cascade Router
Dual-Layer Agentic Memory introduces the non-write, write-new, write-update taxonomy and a cost-aware small-to-large routing cascade that prunes up to 68% redundant external memory while escalating fewer than 50% of inputs.
Periodic Write-Back and Epistemic Shift Analysis
Dual-Layer Agentic Memory uses supervised fine-tuning to internalize external knowledge, achieving 10,683 facts transitioning to non-write and analyzing 1,752 cases of parametric interference in the label transition matrix.
RESULTS
QA EM
86.35%
+-1.44 over Full Store
Tok. F1
0.8846
vs Full Store 0.8980
Store
77.00%
9.00pp less storage than Full Store 100.0%
Routing Cost
Fwd1.7B+39.7% Fwd8B
reduces 8B forward passes compared to 1 × Fwd8B baselines
On the dynamic online ZsRE streaming benchmark, Dual-Layer Agentic Memory’s Write Router traces the EM–storage–compute Pareto frontier. The main result shows that Dual-Layer Agentic Memory retains over 98.3% of Full Store QA EM while substantially reducing external memory and large-model routing compute.
BENCHMARK
On the dynamic online ZsRE streaming benchmark, Dual-Layer Agentic Memory’s Write Router traces the EM–storage–compute Pareto frontier. The main result shows that Dual-Layer Agentic Memory retains over 98.3% of Full Store QA EM while substantially reducing external memory and large-model routing compute.
BENCHMARK
QA EM comparison for base-model policies on the streaming ZsRE benchmark.
KEY INSIGHT
Dual-Layer Agentic Memory’s Write RouterSFT achieves 90.71% QA EM with only 47.85% storage, retaining over 98.2% of the Full Store upper bound at 92.29%.
This is surprising because conventional wisdom suggests exhaustive retention is necessary for high QA accuracy, yet Dual-Layer Agentic Memory shows that aggressive pruning plus consolidation barely reduces EM.
WHY IT MATTERS
Dual-Layer Agentic Memory enables agents to treat memory as a lifecycle, selectively externalizing and later internalizing knowledge while balancing accuracy, storage, and compute.
Builders can now deploy long-horizon LLM agents that avoid unbounded memory growth, maintain near-optimal QA performance, and adapt their epistemic boundaries over time without relying solely on retrieval.
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…