AuthorsLei Wei, Xiao Peng, Xu Dong et al.
TL;DR
FadeMem uses adaptive exponential forgetting in a dual-layer memory with LLM-guided conflict resolution and fusion to reach 29.43 F1 on LoCoMo with 45% storage reduction.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
LLM agents lack selective forgetting, leading to either catastrophic forgetting at context boundaries or information overload within them.
Without adaptive decay, long-term agent memory treats all tokens equally, clogging context windows and degrading multi-hop reasoning over extended interactions.
HOW IT WORKS
FadeMem’s core mechanism combines a Dual-Layer Memory Architecture, Biologically-Inspired Forgetting Curves, Memory Conflict Resolution, and Adaptive Memory Fusion to manage agent memory over time.
Think of FadeMem as a brain-like system where a short-term buffer and a long-term store behave like RAM and disk, with decay curves mimicking hippocampal consolidation.
This design lets FadeMem compress, merge, and forget memories beyond any fixed context window, enabling long-horizon reasoning without exploding storage or losing critical facts.
DIAGRAM
This diagram shows how FadeMem updates, decays, resolves, and fuses memories across days during long-term interactions.
DIAGRAM
This diagram shows how FadeMem is evaluated on MSC, LoCoMo, and LTI-Bench with ablations of its key components.
PROCESS
Dual-Layer Memory Architecture with Differential Forgetting
FadeMem encodes each interaction into memories mi with content, text, strength, timestamps, and access frequency, then assigns them to the Long-term Memory Layer or Short-term Memory Layer using the importance score Ii.
Biologically-Inspired Forgetting Curves
FadeMem applies adaptive exponential decay vi(t) with layer-specific shape parameters and importance-dependent rates, reinforcing memories on access and pruning items when strength falls below ϵprune or exceeds Tmax.
Memory Conflict Resolution
FadeMem retrieves similar memories using cosine similarity, then uses LLM-guided Memory Conflict Resolution to classify relationships as compatible, contradictory, subsumes, or subsumed and update strengths and importance accordingly.
Adaptive Memory Fusion
FadeMem clusters temporally and semantically related memories and runs Adaptive Memory Fusion to merge them into consolidated entries with adjusted decay rates, verified by LLM checks before updating the memory state Mt+Δt.
KEY CONTRIBUTIONS
Dual-layer biologically inspired agent memory with adaptive forgetting
FadeMem introduces a Dual-Layer Memory Architecture with Long-term Memory Layer and Short-term Memory Layer plus Biologically-Inspired Forgetting Curves, achieving different half-lives such as 11.25 and 5.02 days.
Unified framework with LLM-guided conflict resolution and memory fusion
FadeMem uses Memory Conflict Resolution and Adaptive Memory Fusion to enforce temporal consistency and compress redundancy, reaching 68.9% macro-averaged conflict accuracy and 80.4% consistency on LTI-Bench.
Improved multi-hop reasoning with storage savings
FadeMem attains 29.43 multi-hop F1 and 85.9% Factual Consistency Rate on LoCoMo while achieving a Storage Reduction Rate of 0.45, compared to Mem0’s 28.37 F1 and 0.00 SRR.
RESULTS
RP@10
77.2%
+2.4 points over Mem0 on MSC
TCS
0.82
vs Mem0’s 0.79 temporal coherence on MSC
F1
29.43
+1.06 over Mem0 on LoCoMo multi-hop reasoning
SRR
0.45
45% storage reduction on LoCoMo compared to 0.00 for Mem0 and LangChain
On MSC and LoCoMo, which test conversational memory and very long-context multi-hop reasoning, FadeMem shows higher retrieval precision, temporal consistency, and F1 while cutting storage by 45% compared to Mem0 and MemGPT. These results demonstrate that FadeMem’s adaptive forgetting preserves critical information without keeping redundant or outdated memories.
BENCHMARK
On MSC and LoCoMo, which test conversational memory and very long-context multi-hop reasoning, FadeMem shows higher retrieval precision, temporal consistency, and F1 while cutting storage by 45% compared to Mem0 and MemGPT. These results demonstrate that FadeMem’s adaptive forgetting preserves critical information without keeping redundant or outdated memories.
BENCHMARK
Multi-hop F1 on LoCoMo for different memory systems.
KEY INSIGHT
FadeMem retains 82.1% of critical facts on LTI-Bench while using only 55.0% storage, and still reaches 85.9% Factual Consistency Rate.
This is surprising because aggressive forgetting is usually assumed to hurt accuracy, yet FadeMem’s adaptive decay and fusion improve multi-hop F1 to 29.43 while discarding 45% of stored memories.
WHY IT MATTERS
FadeMem unlocks long-horizon agents that can maintain coherent, temporally consistent memories over 30-day interactions without exploding storage or context windows.
Builders can now design agents that continuously learn, update, and compress user-specific knowledge, using FadeMem to keep only what matters while still supporting complex multi-hop reasoning.
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…