Source-linked AI summary
PraMem: Practice-derived Experiential Memory for Long-horizon Behavior Prediction
Zhuoqun Li, Boxi Cao, Jiawei Chen, Hanshu Zhou, Ruoxi Xu, Guiping Jiang, Ruotong Pan, Tingting Gao, Han Li, Xiangyu Wu, Hongyu Lin, Yaojie Lu, Xianpei Han, Le Sun
TL;DR
Long-horizon behavior prediction remains difficult for LLMs when behavioral patterns must be induced from lengthy histories. PraMem instead practices over those histories to build experiential memory, achieving superior performance across diverse tasks, including ACC improving from 73.5 (Truncation) to 84.7 on OmniBehavior.
Problem
Long-horizon behavior prediction remains challenging because LLMs must induce behavioral patterns from sequences spanning thousands of steps.
Method
PraMem is a training-free framework that iteratively practices on historical sequences to build pattern and bias-alert experiential memories for prediction.
Results
Across diverse long-horizon behavior prediction tasks, PraMem achieves superior performance, improving ACC from 73.5 (Truncation) to 84.7 on OmniBehavior.
Takeaways & Limitations
Experiential memory remains effective across different LLM backbones, offering practical value because it can be constructed once and reused.
Takeaways & Limitations
PraMem’s iterative practice process requires more time to construct experiential memory than conventional extraction, aggregation, and retrieval methods.
Abstract
from arXiv · showhide
Long-horizon behavior prediction aims to infer a user's next action based on a lengthy historical sequence, playing a crucial role in artificial intelligence field. The rise of large language models (LLMs) offers a promising direction for sequential behavior prediction, yet LLMs struggle with latent behavioral pattern induction and model-intrinsic cognitive biases when tackling long-horizon behavior prediction. Prior memory management methods follow a context-compression paradigm that attempts to address this task by alleviating the historical sequence burden, yet fail to resolve the core challenges. In this paper, we advocate a paradigm shift that reframes the lengthy historical sequence from a burden into a valuable resource to be exploited, and accordingly propose PraMem, which conducts beforehand practice over the lengthy historical sequence to build an experiential memory, thereby serving as the assisted input for accurate long-horizon behavior prediction. Extensive experiments across diverse tasks demonstrate that PraMem achieves superior performance than prior methods, and more in-depth analyses provide valuable insights into the mechanism and evolution of the experiential memory. Code: https://github.com/icip-cas/PraMem.
1 Introduction
PraMem reframes lengthy historical sequences from a context burden into a resource by practicing over them to build experiential memory for long-horizon behavior prediction. Its training-free framework combines pattern and bias-alert experiences with iterative maintenance and self-review to improve reliability.
- Problem: Long-horizon behavior prediction infers a user’s current action from behavioral patterns in lengthy historical scene-action sequences.The task supports recommendation systems, cognitive science, and interactive artificial intelligence.
- PraMem: PraMem iteratively performs experience trial, reflective proposal generation, and consensual experience adjustment to dynamically maintain its experiential memory.This process maintains both pattern experience and bias-alert experience before using them for prediction.
- Motivation: Prior memory-management methods compress historical context by extracting, aggregating, and retrieving relevant behavioral records, but this paradigm fails to effectively address core challenges.The cited passage frames compression as alleviating the long-context burden rather than exploiting the sequence as a resource.
- PraMem: PraMem is a training-free framework that practices beforehand over lengthy histories to build a time-evolving experiential memory for accurate prediction.The memory contains pattern experience for presenting behavioral patterns and bias-alert experience for alerting LLM intrinsic biases.
- Self-review: PraMem’s self-review mechanism filters reflective proposals lacking sufficient groundedness or generalizability, supporting the framework’s overall effectiveness.It perturbs practice samples to test whether proposals remain valid when underlying behavioral patterns change significantly.
- Evaluation: Extensive experiments across diverse long-horizon behavior prediction tasks show superior performance over strong memory-management baselines, while ablations verify component and self-review effectiveness.Further analyses examine the evolution of experiential memory and provide insights into PraMem’s mechanism.
2 PraMem
PraMem shifts long-horizon behavior prediction from directly processing lengthy histories to beforehand practice that builds a user-specific, time-evolving experiential memory. It combines behavioral-pattern guidance, bias alerts, and iterative trial, reflection, self-review, and consensus-based adjustment.
- PraMem Process: PraMem practices over the lengthy historical sequence to build a user-specific, time-evolving experiential memory that assists accurate prediction.The memory is additional input to the LLM during prediction.
- Vanilla Process: Vanilla prediction autoregressively conditions the LLM on the textual concatenation of the historical sequence H_t and current scene s_t.The goal is to capture underlying behavioral patterns without bias.
- Experiential Memory: The experiential memory contains Pattern Experience for user behaviors and Bias-alert Experience for intrinsic biases likely to arise during prediction.Examples include product preferences, monthly purchasing patterns, pre-purchase favorites, and avoiding proximal bias.
- Existing Experience Trial: PraMem samples q historical segments, prompts deep-thinking practice under the current memory, and preserves temporal order using target records later than the memory state.Each practice sample combines a short preceding suffix with a target scene and ground-truth action.
- Reflective Proposal Generation: After each trial, reflection generates revision, supplementation, or pruning proposals, while SelfReview retains proposals that are grounded and generalizable.Reliable proposals accumulate in a pool for later experience adjustment.
- Consensual Experience Adjustment: Every T rounds, PraMem performs consensus-driven memory adjustment using consistently supported proposal operations, then removes adopted proposals and retains the rest.Periodic adjustment avoids instability from occasional user behaviors.
3 Self-review Mechanism
PraMem’s experiential memory depends on reflective proposals that must be grounded in practice samples and sufficiently general. The self-review mechanism tests these properties through perturbed historical sequences and virtual prediction scenes.
- 3 Self-review Mechanism: Reflective proposals govern experiential-memory generation and evolution, so their groundedness and generalizability critically determine PraMem’s effectiveness.They should capture authentic predictive requirements without being fabricated or overly sample-specific.
- Groundedness Review: Groundedness review perturbs a sample’s historical sequence to alter behavioral patterns, then checks whether the original proposal still holds.A proposal that remains valid after perturbation is treated as insufficiently grounded.
- Generalizability Review: Generalizability review rewrites a sample’s prediction scene into same-category virtual scenes, then tests whether the proposal unambiguously selects the true scene.If it does, the proposal is considered overly specific to the current practice sample.
4 Experiments
Experiments across OminiBehavior and MovieLens-1M show that PraMem substantially outperforms memory-management baselines, transfers across LLM backbones, and improves through iterative practice. Ablations further support the contributions of its twofold experience, self-review, and practice mechanisms.
- Experimental Setup: The evaluation uses OminiBehavior’s four-scenario long-horizon tasks and MovieLens-1M’s 5-class rating prediction from users with more than 300 records.MovieLens-1M results are compared using confusion matrices and diagonal dominance.
- Main Results: PraMem achieves the best overall performance on OminiBehavior, improving ACC from 73.5 with Truncation to 84.7 and F1 from 24.7 to 31.6.Mem0, an advanced memory-management baseline, achieves ACC of 74.7.
- Main Results: The practice-derived experiential memory transfers across backbones, yielding substantial gains with Qwen3.5-35B-A3B as well as GPT-OSS-120B.Only GPT-OSS-120B is used to construct the memory, while both backbones are used for evaluation.
- Experiential Memory Evolution: Performance rises with practice rounds, reaching 79.0 ACC and 28.3 F1 at round 10, 83.5 ACC and 30.7 F1 at round 40, and 84.7 ACC and 31.6 F1 at round 80.The experiential-memory length increases through round 60 and then stabilizes, while performance continues upward.
- Ablation Studies: Removing deep-thinking trials, reflective proposals, consensual adjustment, either experience type, or either self-review component reduces performance.The tested components are pattern experience, bias-alert experience, groundedness review, and generalizability review.
5 Related Work
Related work studies experiential memory as a mechanism for agent self-evolution and memory management for long-horizon behavior prediction. Experiential memory reuses encoded interaction experiences, while prior behavior-prediction methods mainly reduce long-context burden through extraction, summarization, and retrieval.
- Experiential Memory: Experiential memory encodes historical interactions, successful experiences, and failure lessons into durable, retrievable representations for later reuse.This reuse helps agents avoid repeating mistakes and improve decision quality.
- Memory for Behavior Prediction: Prior behavior-prediction methods manage historical sequences through extraction, summarization, and retrieval to alleviate LLMs’ long-context burden.The cited works focus on memory management over historical sequences for long-horizon behavior prediction.
- Memory for Behavior Prediction: These memory-management approaches do not directly address the two core challenges of long-horizon behavior prediction.The passage contrasts burden alleviation with directly addressing the task’s underlying challenges.
6 Discussion
PraMem reframes lengthy historical sequences as resources by practicing over them beforehand to construct experiential memory for long-horizon behavior prediction. Its main limitation is the higher offline memory-construction cost compared with conventional memory management methods.
- Conclusion: PraMem builds experiential memory through iterative existing-experience trials, reflective proposal generation, and consensual experience adjustment.This maintains pattern experience for presenting user behavioral patterns and bias-alert experience for al…
- Limitations: PraMem requires more time for memory construction than conventional extraction-, aggregation-, and retrieval-based methods, although the process is conducted offline.The offline process does not affect actual prediction efficiency, but remains a non-trivial limitation.
A Prompt for Existing Experience Trial
The existing experience trial prompt guides the LLM to practice on historical-sequence samples using accumulated experiential memory, while explicitly identifying relevant pattern experience.
- Existing Experience Trial: The prompt is used during PraMem’s existing experience trial, where the LLM practices beforehand on constructed samples from the historical sequence.This stage turns historical behavior into practice tasks for experience accumulation.
- Existing Experience Trial: For each practice task, the LLM predicts the action using both the historical behavior sequence and accumulated experience.The prediction is conditioned on the current experiential memory.
- Existing Experience Trial: The prompt separately asks the LLM to identify relevant pattern experience, making experiential-memory use explicit and controllable.This separates experience identification from the action-prediction process.
B Prompt for Reflective Proposal Generation
The reflective proposal generation prompt has the LLM compare predictions with ground-truth labels after each practice round. It converts trial-and-error outcomes into experiential-memory revisions while identifying behavioral patterns and user-specific deficiencies.
- Reflective Proposal Generation: The prompt is used after each completed practice round to compare the LLM’s predictions with ground-truth labels.This comparison forms the basis for reflection.
- Reflective Proposal Generation: It transforms trial-and-error outcomes into candidate revisions for experiential memory.These revisions support gradual improvement across practice rounds.
- Reflective Proposal Generation: The LLM analyzes both correct and incorrect predictions to induce behavioral patterns and identify deficiencies for the current user.The prompt targets reliable pattern induction and user-specific weakness detection.
C Prompt for Consensual Experience Adjustment · D Prompt for Groundedness Review · E Prompt for Generalizability Review
PraMem uses three review prompts to stabilize experiential memory and evaluate reflective proposals. The prompts address consensus, grounding in historical behavior, and transfer to similar prediction scenarios.
- C Prompt for Consensual Experience Adjustment: The consensual adjustment prompt periodically reviews accumulated proposals, grouping similar ones and adopting viewpoints supported by multiple proposals.This prevents unstable memory changes caused by occasional behaviors or noisy reflections.
- D Prompt for Groundedness Review: The groundedness review tests whether reflective proposals derive genuinely from the historical sequence rather than being generic or hallucinated statements.It first asks the LLM to design perturbation plans that substantially change preference signals in the historical behavior sequence.
- E Prompt for Generalizability Review: The generalizability review tests whether reflective proposals capture transferable experience rather than merely memorizing one practice sample.It generates virtual scenes matching the real prediction scene’s broad type and format while differing in key entities or details, then evaluates each proposal.
F Case Study
The case study compares PraMem’s experiential memory with MemOS for the same user, showing that PraMem stores condensed behavioral patterns and bias alerts rather than only historical facts.
- Case Study: PraMem and MemOS are compared on the memory content maintained for the same user.The comparison is presented as a case study illustrating PraMem’s advantages over traditional memory management.
- Case Study: PraMem’s distilled memory contains pattern experiences that concisely characterize the user’s behavioral regularities.Examples concern preferences for authors with social or emotional tags and viewing duration determined by content type, title keywords, and video length.
- Case Study: PraMem also includes bias-alert experiences that explicitly flag cognitive biases to which the LLM tends to fall.
- Case Study: MemOS retains factual records about what happened, whereas PraMem constructs experience-oriented memory.Compressed and retrieved baseline records still require the LLM to induce behavioral patterns from fragmented facts during prediction and do not address intrinsic cognitive biases.
G Efficiency Report
PraMem incurs higher memory-construction overhead because it performs extensive iterative practice over historical sequences. This cost can be handled offline, avoiding added online-prediction latency while trading moderately higher construction cost for performance gains.
- Construction overhead: PraMem’s iterative practice over historical sequences leads to higher average per-user memory-construction time than conventional compression-based methods.Table 3 reports the average per-user construction cost for PraMem and three advanced memory-based baselines.
- Deployment efficiency: Memory construction can be conducted offline in practical applications, so it does not increase online-prediction latency.This allows the construction overhead to be separated from online inference.
- Efficiency trade-off: PraMem trades a moderately larger offline construction cost for substantial performance gains, which the paper considers acceptable.The paper discusses this efficiency trade-off further in its Limitations section.
H Potential Negative Societal Impact
PraMem’s deployment raises privacy concerns because it builds experiential memory from lengthy user behavior sequences. These sequences may contain sensitive preferences and interaction histories, creating a risk of unintended behavioral-trace exposure without proper data protection.
- PraMem relies on lengthy user behavior sequences to build experiential memory.
- These sequences may contain sensitive personal information, including user preferences and interaction histories.
- Without proper data protection mechanisms, deployment risks unintended exposure of user behavioral traces.