cs.AIApr 2, 2026

Novel Memory Forgetting Techniques for Autonomous AI Agents: Balancing Relevance and Efficiency

Authors: Payal Fofadiya, Sunil Tiwari

Organizations: Fulloop

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Long-horizon conversational agents require persistent memory for coherent reasoning, yet uncontrolled accumulation causes temporal decay and false memory propagation. Benchmarks such as LOCOMO and LOCCO report performance degradation from 0.455 to 0.05 across stages, while MultiWOZ shows 78.2% accuracy with 6.8% false memory rate under persistent retention. This work introduces an adaptive budgeted forgetting framework that regulates memory through relevanceguided scoring and bounded optimization. The approach integrates recency, frequency, and semantic alignment to maintain stability under constrained context. Comparative analysis demonstrates improved long-horizon F1 beyond 0.583 baseline levels, higher retention consistency, and reduced false memory behavior without increasing context usage. These findings confirm that structured forgetting preserves reasoning performance while preventing unbounded memory growth in extended conversational settings.

Figures & tables

RefDataset UsedMethodologyLimitationEvaluation Results (Reported in Paper)
Honda et al. [ 11 ]Dialogue simulation experimentsACT-R based activation model with temporal decay, frequency reinforcement, semantic similarity spreading, and Gaussian noiseNo comparative benchmark evaluation; no memory–performance tradeoff experimentsSimulation experiments validated reinforcement through repeated topics and stochastic variability; optimal decay parameters identified for contextual sensitivity and stability
Ming et al. [ 12 ]Custom dialogue datasetLong-term + short-term integration, indexing-based retrieval modification, caching prefetch inspired by forgetting curveNo explicit deletion policy; no memory budget constraint analysisAnswer Accuracy 88.4%; Retrieve Accuracy 0.938; Recall Accuracy 0.663; Coherence 0.948; Execution time reduced 21.45% (5.17s →\rightarrow 4.06s); Hit rate 35.35%
Xiao et al. [ 13 ]∞\infty -Bench, LongBenchBlock-level context memory; sliding window + selective memory lookup; CPU-GPU offloadingNo controlled forgetting strategy; no comparative deletion experimentsProcesses up to 1024K tokens; 34% reduction in time consumption; 34% GPU memory usage compared to full attention
Shen et al. [ 14 ]LongBench; Needle-In-A-Haystack; Ruler; InfiniteBenchLayer-wise KV cache eviction with dynamic head and layer budget allocation via residual information loss minimizationDesigned for KV cache compression; does not model persistent long-term memory across sessionsLongBench Avg: 41.45 (25% KV budget), 41.43 (12.5%), 41.05 (6.25%); maintains stable performance under aggressive compression; 9 ×\times faster decoding at 128K tokens
Saleh et al. [ 15 ]Multi-agent pub/sub experimental setupDistributed memory management with intent-indexed bipartite graph supporting storage, retrieval, deletion, and updateFocused on systems efficiency; no cognitive forgetting modelExample (o3-pro): Storage 82.58s, 2.10% CPU, 3.70% memory (vs Baseline1 132.22s, 8.30%, 4.70%). Retrieval 67.93s, 2.00% CPU, 3.60%. Deletion 96.07s, 2.10% CPU, 3.70%. Update 90.77s, 2.00% CPU, 3.70%.
Mirani et al. [ 16 ]GSM8k, AQuA, BBH, LongBenchStructured KV cache compression using expander graphsTargets inference compression; not long-term agent memory controlKV size 32.1%; GSM8k 55.04; AQuA 35.04; BBH 53.45; FP16 baseline (100% KV) GSM8k 54.21; AQuA 38.19; BBH 53.66
TABLE I: Comparison of Multi-Layer Memory and Long-Context Management Approaches
Fig. 1: Architecture of the Adaptive Budgeted Forgetting Framework. The system integrates multi-layer memory organization, adaptive relevance-guided control, and budget-constrained selection to maintain long-horizon consistency under fixed memory constraints.
 Mt←Fstore(Mt−1,ot,at)\mathcal{M}_{t}\leftarrow\mathcal{F}_{store}(\mathcal{M}_{t-1},o_{t},a_{t}) 
Algorithm 1 Adaptive Budgeted Forgetting via Constrained Relevance Maximization
DatasetMetricReported Result
LOCOMO [ 18 ]Overall F151.6 (gpt-4-turbo); 31.2 (gpt-3.5); 18.7 (Mistral-7B)
LOCCO [ 17 ]Memory Score0.455 →\rightarrow 0.05 (Openchat-3.5); 48.25% retained (ChatGLM3-6B)
MultiWOZ 2.4 [ 23 ]Accuracy / FMR78.2% Accuracy; 6.8% False Memory Rate
TABLE II: Reported Benchmark Results from Prior Studies
Fig. 2: Benchmark results across LOCOMO, LOCCO, and MultiWOZ 2.4 highlighting long-horizon reasoning performance, memory decay behavior, and false memory rates.
RefACCUPreciRecallF1Metrics
[ 12 ]88.4%0.9380.663–Coherence 0.948; Exec time ↓21.45%; Hit rate 35.35%
[ 21 ]93.3%91.2%–26.53Coherence 92.3%; LoCoMo Rank 1.0; +12.56% Single-hop; +14.66% Temporal
[ 22 ]–––0.583BLEU-1 0.599; Entity F1 0.719; Multi-hop 0.550
[ 23 ]78.2%––0.654FMR 6.8%; DAR 82.4%; CDR 41.2%
[ 16 ]55.04%–––KV size 32.1%; AQuA 35.04; BBH 53.45
[ 25 ]79.66%–75.33%–τdel=0.9\tau_{del}=0.9 : A=81.46; Fk=88.30; Mnemonic loss S 74.02
TABLE III: Comparison of Performance Metrics with Prior Work
Fig. 3: Proposed method across key performance metrics, with solid-line overlay highlighting relative improvement trends.
StudyBudget ReductionPerformance ChangeEfficiency Gain
[ 14 ]KV 25% →\rightarrow 6.25%LongBench 41.45 →\rightarrow 41.059 ×\times faster decoding (128K tokens)
[ 20 ]Context 100% →\rightarrow 64.98%SR 21.00 →\rightarrow 42.00; PR 38.61 →\rightarrow 62.55Time 100% →\rightarrow 80.58%
[ 23 ]Tiered write filteringMultiWOZ Acc 78.2%; FMR 6.8%Reduced false memory accumulation
[ 17 ]Temporal decay (T1 →\rightarrow T6)M1M_{1} 0.455 →\rightarrow 0.05Consistency model 98% accuracy
TABLE IV: Long-Horizon Performance Under Memory Constraints (Reported Results)
Fig. 4: LOCOMO F1, LOCCO retention, MultiWOZ accuracy, and False Memory Rate (FMR) for prior work and the proposed framework.
StudySettingRetention / AccuracyFalse Memory Metric
[ 17 ]LOCCO (T1 →\rightarrow T6)M1M_{1} : 0.455 →\rightarrow 0.05Consistency model: 98%
[ 23 ]MultiWOZ 2.4Accuracy: 78.2%FMR: 6.8%
TABLE V: False Memory and Retention Stability in Prior Benchmarks
More in this topic

Agent Memory Management

Agent memory management studies how AI agents store, update, retrieve, compress, and forget information across long-running tasks.

See all 446 papers

Explore similar work

CardsList
Apr 21, 2026

From Recall to Forgetting: Benchmarking Long-Term Memory for Personalized Agents

Personalized agents that interact with users over long periods must maintain persistent memory across sessions and update it as circumstances change. However, existing benchmarks predominantly frame long-term memory evaluation as fact retrieval from past conversations, providing limited insight into agents' ability to consolidate memory over time or handle frequent knowledge updates. We introduce Memora, a long-term memory benchmark spanning weeks to months long user conversations. The benchmark evaluates three memory-grounded tasks: remembering, reasoning, and recommending. To ensure data quality, we employ automated memory-grounding checks and human evaluation. We further introduce Forgetting-Aware Memory Accuracy (FAMA), a metric that penalizes reliance on obsolete or invalidated memory when evaluating long-term memory. Evaluations of four LLMs and six memory agents reveal frequent reuse of invalid memories and failures to reconcile evolving memories. Memory agents offer marginal improvements, exposing shortcomings in long-term memory for personalized agents.

Md Nayem Uddin, Kumar Shubham, Eduardo Blanco +2

Arizona State University · Genies · University of Arizona

Long-Horizon Agent EvaluationAI Agent BenchmarksMemory Agent BenchmarksLong-Term Conversational MemoryAgent Memory Management
AskPDF
Mar 31, 2026

Multi-Layered Memory Architectures for LLM Agents: An Experimental Evaluation of Long-Term Context Retention

Long-horizon dialogue systems suffer from semanticdrift and unstable memory retention across extended sessions. This paper presents a Multi-Layer Memory Framework that decomposes dialogue history into working, episodic, and semantic layers with adaptive retrieval gating and retention regularization. The architecture controls cross-session drift while maintaining bounded context growth and computational efficiency. Experiments on LOCOMO, LOCCO, and LoCoMo show improved performance, achieving 46.85 Success Rate, 0.618 overall F1 with 0.594 multi-hop F1, and 56.90% six-period retention while reducing false memory rate to 5.1% and context usage to 58.40%. Results confirm enhanced long-term retention and reasoning stability under constrained context budgets.

Sunil Tiwari, Payal Fofadiya

Fulloop

Persistent Memory for Language ModelsLong-Term Conversational MemoryAgent Memory ManagementHierarchical MemoryLong-Horizon Agent Evaluation
AskPDF
Mar 27, 2026

Cognis: Context-Aware Memory for Conversational AI Agents

LLM agents lack persistent memory, causing conversations to reset each session and preventing personalization over time. We present Lyzr Cognis, a unified memory architecture for conversational AI agents that addresses this limitation through a multi-stage retrieval pipeline. Cognis combines a dual-store backend pairing OpenSearch BM25 keyword matching with Matryoshka vector similarity search, fused via Reciprocal Rank Fusion. Its context-aware ingestion pipeline retrieves existing memories before extraction, enabling intelligent version tracking that preserves full memory history while keeping the store consistent. Temporal boosting enhances time-sensitive queries, and a BGE-2 cross-encoder reranker refines final result quality. We evaluate Cognis on two independent benchmarks -- LoCoMo and LongMemEval -- across eight answer generation models, demonstrating state-of-the-art performance on both. The system is open-source and deployed in production serving conversational AI applications.

Parshva Daftari, Khush Patel, Shreyas Kapale +2

Lyzr Research

Agent Memory ManagementMemory-Augmented Language Model AgentsLong-Term Conversational MemoryHybrid IR
AskPDF