Continual Personalization of Large Language Models: Memory, Parameters, and the Problem of Forgetting
A structured survey of where user knowledge is stored, how personalization updates over time, and what signals drive those updates.
Reviews and research notes on the ideas shaping intelligent systems. Browse an entry below to read its complete text or download the original document.
A structured survey of where user knowledge is stored, how personalization updates over time, and what signals drive those updates.
Authors: Tejas S. Vaij, Dr. Sunil B. Mane · Structured survey, working draft
Continual Personalization of Large Language Models: Memory, Parameters, and the Problem of Forgetting
Structured survey, working draft. Entries marked † in the reference list were not checked against the primary source while drafting and must be verified before submission. Text in [square brackets] is a placeholder to fill in after the search protocol in Section 3 is run.
Large language models (LLMs) are general-purpose, yet most users want an assistant that knows their preferences, goals, and working style and keeps learning them. Existing personalization work mostly stores user knowledge outside the model, in prompts, profiles, retrieval indices, or conversational memory. A smaller and newer line of work writes user knowledge into parameters, typically through per-user parameter-efficient fine-tuning (PEFT), and a still smaller line studies how to keep such adaptation going over time without forgetting. This survey organizes the literature along three axes: where user knowledge is stored, how it is updated, and what signal drives the update. We compare four architectures (base model, memory/retrieval, parameter-level, and hybrid), review continual-learning techniques relevant to per-user adaptation, and examine how personalization is evaluated. We find that the parametric and hybrid directions are active but young, that long-horizon evaluation on conversational feedback streams is thin, and that the separation of stable preferences from transient instructions has received little dedicated treatment in the sources reviewed so far. We close with open problems and an illustrative, unevaluated architecture for local continual personalization. [Final abstract numbers, such as papers screened and included, to be added after the search.]
Keywords: LLM personalization, continual learning, LoRA, long-term memory, retrieval-augmented generation, catastrophic forgetting, on-device learning
Pretrained LLMs behave identically for every user unless something user-specific is supplied. The dominant way to supply it is external: place a profile or retrieved history in the prompt, or maintain a memory store that is queried at inference time. These approaches are simple and inspectable, but the model itself learns nothing about the user. A different approach is to adapt the model's parameters to the user, as in One PEFT Per User (OPPU) (Tan et al., 2024). A user can then own a personal adapter, and user knowledge no longer has to compete for space in the context window.
The question driving this survey is whether, and how, an LLM can keep internalizing an individual user over extended interaction, and how this compares with memory and retrieval. We make three contributions:
A taxonomy that treats storage location, update regime, and update signal as separate axes, since common categories such as hybrid and feedback-driven personalization overlap.
A structured comparison of four architectures (base, memory/RAG, parametric, hybrid) and a review of what existing experiments have and have not compared.
A gap analysis separating problems that existing work already addresses from problems that appear open.
Related surveys. Zhang et al. (2025) survey LLM personalization with taxonomies for personalization granularity (user, persona, global), technique, and evaluation. Liu et al. (2025) organize the area into prompting for personalized context, fine-tuning for personalized adapters, and alignment for personalized preferences, and note that lifelong updating is under-explored in current benchmarks. In parallel, Shi et al. (2024) and Yang et al. (2024) survey continual learning for LLMs. The present survey differs by centering the interaction between these literatures: personalization under continual updates, with the memory-versus-parameters trade-off as the organizing question.
Setting. A user u interacts with an assistant over time. At step t the assistant observes a query x_t, produces a response y_t, and may observe feedback f_t (explicit correction, approval, rejection, or implicit behavior). The interaction stream is D_u = {(x_t, y_t, f_t)}.
Components. The assistant is a base model with parameters θ, optionally extended by (i) a user-specific adapter φ_u, and (ii) an external memory M_u with a retrieval function R. A response is drawn from p(y | x, R(M_u, x); θ + φ_u).
Four architectures.
A, Base: φ_u and M_u both empty.
B, Memory/RAG: φ_u empty, M_u non-empty.
C, Parametric: φ_u non-empty, M_u empty.
D, Hybrid: both non-empty.
Continual personalization. φ_u is updated over time from φ_u(t−1) using new data D_u(t), under at least three desiderata: (a) responsiveness to genuine preference change, (b) stability of previously learned user knowledge, and (c) retention of the base model's general capabilities. Any method must also decide which interactions are worth learning from.
Granularity. Following Zhang et al. (2025), personalization can be user-level (finest, needs per-user data), persona-level (users grouped into personas), or global (population preferences). This survey focuses on user-level personalization, with group-level sharing considered where it reduces data or storage needs.
This is a structured survey: it follows a documented, reproducible search protocol, but it does not claim the exhaustive coverage or dual-reviewer screening of a full systematic review. That choice reflects a fast-moving field in which many relevant papers are preprints.
Sources. arXiv, ACL Anthology, Semantic Scholar, OpenReview, ACM Digital Library, IEEE Xplore, and Google Scholar, plus forward and backward citation snowballing from the seed set in Section 3.3. Search date: [date]. Time window: 2020 to [search date], with earlier foundational work admitted as background.
Query families (adapted to each database's syntax):
(personalized OR personalization) AND (LLM OR large language model) AND (LoRA OR adapter OR parameter-efficient)
(continual OR lifelong OR online) AND personalization AND (LLM OR language model)
(long-term memory OR conversational memory) AND (LLM OR chat assistant) AND (personali* OR user)
(user feedback OR user corrections) AND (preference optimization OR DPO OR reward model) AND personali*
(preference drift OR behavior shift) AND user AND LLM
(on-device OR local) AND (personalization OR fine-tuning) AND LLM
Inclusion criteria. A work is included if it (1) adapts an LLM or an LLM-based system to an individual user or small group, (2) contributes a method, benchmark, or empirical study, (3) falls in the time window, and (4) is in English.
Exclusion criteria. Recommender-system work with no language-model component, persona prompting that uses no user data, opinion pieces without analysis, duplicate versions of one work, and papers too thin to classify.
Screening. Records are deduplicated, screened by title and abstract, then by full text. Counts per stage: identified [N], after deduplication [N], full-text assessed [N], included [N]. Each included work is tagged peer-reviewed or preprint, and preprint status is re-checked before submission.
Seed set. The surveys above; LaMP (Salemi et al., 2024); OPPU (Tan et al., 2024); LongMemEval (Wu et al., 2025); SPRInG (Kim and Kim, 2026).
Existing categories (prompt-based, profile-based, memory-based, RAG-based, adapter-based, fine-tuning-based, continual, feedback-driven, hybrid) are not mutually exclusive. We therefore classify each work along three axes.
| Axis | Values |
|---|---|
| Storage location | In context (prompt, profile) · external store (memory, retrieval index, knowledge graph) · parameters (adapter, full fine-tune) · more than one |
| Update regime | Static or offline · periodic batch · online or continual |
| Update signal | Behavioral history · explicit feedback or corrections · preference pairs · self-generated or model-estimated signals |
Two cross-cutting dimensions apply to every cell: cost (training compute, inference latency, storage) and privacy (where data and adapters reside, and what can be extracted from them).
Mapping of common categories. Prompt-based and profile-based methods sit in the context row. Memory-based and RAG-based methods sit in the external-store row and differ mainly in how memory is written, indexed, and read. Adapter-based and fine-tuning-based methods sit in the parameters row. Continual personalization is a value on the update axis. Feedback-driven personalization is a value on the signal axis. Hybrid personalization spans storage locations.
Mechanism. Raw history, summaries, or extracted facts are stored outside the model and retrieved at inference time. LaMP (Salemi et al., 2024) is the most-used benchmark for this family: seven personalized tasks (three classification, four generation) with both user-based and time-based splits, plus retrieval-augmented baselines in which retrieved profile items are placed in the prompt. Their results show that profile augmentation improves over models that ignore the profile.
Conversational memory. For assistants, the question shifts from retrieving a user's documents to maintaining a memory across sessions. LongMemEval (Wu et al., 2025) evaluates five abilities: information extraction, multi-session reasoning, temporal reasoning, knowledge updates, and abstention. It reports a substantial accuracy drop for commercial assistants and long-context models on sustained interactions. It also decomposes memory systems into indexing, retrieval, and reading stages and finds that storing history at the level of rounds works better than whole sessions. LoCoMo (Maharana et al., 2024†) is a companion benchmark of long conversations with questions spanning single-hop, multi-hop, temporal, commonsense, and adversarial types.
Profiles and structured stores. Westhäußer et al. (2025) combine persistent memory with dynamically maintained user profiles in an agent and report a five-day pilot study of perceived personalization. Knowledge Graph Tuning (KGT†) extracts factual triples from queries and feedback and updates a knowledge graph instead of model weights.
Strengths. Immediate updates, inspectable and editable content, easy deletion, no training cost, and no forgetting of base capabilities.
Limitations. Retrieval quality bounds personalization quality. Costs grow with history. Recency signals are weak in large stores. Everything must pass through the context window, and the model never internalizes patterns that are implicit rather than stated. Kim and Kim (2026) make a related point, noting that non-parametric methods struggle as histories expand.
Mechanism. User data trains a small set of parameters (low-rank adapters such as LoRA, Hu et al., 2022†, or its quantized variant QLoRA, Dettmers et al., 2023†) while the base model stays frozen.
One PEFT per user. OPPU (Tan et al., 2024) trains a personal PEFT module per user and plugs it into a shared base. The authors motivate it by user ownership and privacy of the model and by better handling of behavior shift. They report that it outperforms prompt-based methods across the seven LaMP tasks, and they show it can be combined with retrieval and profile augmentation.
Scaling beyond one adapter per user. Liu et al. (2025) identify the storage and compute burden of one adapter per user and the lack of cross-user sharing as limitations, and review hierarchical designs, such as grouping users and mixing population-, group-, and user-level LoRA experts (PROPER†). Other recent work replaces per-user adapters with generated or compact representations: hypernetworks that map a user profile to adapter weights, and prefix-state methods such as TAP-PER (2026 preprint†). PEFT-U† provides a benchmark for user-level PEFT personalization.
Strengths. The model can capture implicit style and patterns, inference needs no retrieval step, and the adapter is portable and user-owned.
Limitations. Training cost and latency before knowledge takes effect, difficulty of editing or deleting a single fact, a risk of memorizing sensitive content, sparse per-user data, and for continual use, forgetting (Section 7).
Continual learning for LLMs. Shi et al. (2024) distinguish vertical continuity (general to specific) from horizontal continuity (across time and domains) and survey continual pre-training, domain-adaptive pre-training, and continual fine-tuning, including continual instruction tuning, model refinement, and alignment. Yang et al. (2024) split the field into offline and online continual learning and into traditional, PEFT-based, instruction-tuning-based, and continual pre-training methods. The core problem, catastrophic forgetting, goes back to work such as elastic weight consolidation (Kirkpatrick et al., 2017†).
What is different about personalization. Interaction streams are continuous and noisy, preferences drift gradually rather than switching in discrete steps, data per user is small, and the cost of an unwanted behavior change is borne by a single person. Generic continual-learning methods (replay, regularization, orthogonal subspaces) therefore need adaptation.
Continual personalization methods. SPRInG (Kim and Kim, 2026) is, among the works examined, the clearest instance. In training it scores interactions by novelty relative to the current user adapter and fine-tunes only on those indicating genuine preference change, while a replay buffer keeps hard-to-learn historical examples. At inference it gates retrieval from that buffer and fuses retrieved and parametric predictions by logit interpolation. The authors evaluate on long-form tasks from LongLaMP and compare with continual-learning baselines (including O-LoRA-, EWC-LoRA-, and replay-style methods) and standard RAG; results are as reported by the authors in a January 2026 preprint and have not been independently replicated. A further preprint, COPE†, addresses continual personalization under sparse user feedback using user embeddings and self-evaluation; we have only reviewed its title and reference list and do not characterize its results here. PersonalLLM† builds simulated users from reward models and is presented as a testbed for settings where a user's feedback is sparse over time.
Gaps within this family. Evaluations so far use document-authoring benchmarks rather than assistant conversations, report limited retention of general capabilities, and rarely compare against a well-tuned memory-only system.
Signals. Explicit corrections, approvals and rejections, edits, choices between alternatives, and implicit behavior can all serve as training signals, but they differ in reliability, volume, and ambiguity.
From feedback to training data. Preference-optimization methods (RLHF and direct-preference variants) learn from comparisons, and personalized versions condition on or train per-user reward signals. Simulated-user testbeds such as PersonalLLM† exist, but real-user, long-horizon feedback datasets remain scarce in the sources reviewed.
Stable preferences versus transient instructions. A user saying 'be brief' may mean always, or only now. We did not identify, in the sources reviewed so far, a dedicated treatment that classifies conversational statements as durable preference, task-local instruction, or noise before they reach training. Related ideas appear in SPRInG's novelty-based selection, which judges interactions by how surprising they are to the current adapter, but novelty is not the same as durability. Memory systems handle part of this problem through write, update, and supersede operations, and LongMemEval's knowledge-update tasks test it for stored facts, not for weights. [Confirm with targeted search before asserting a gap.]
Contradiction and change. Memory stores can overwrite an old fact. Weights cannot be edited as cleanly, so superseded preferences may persist in a user adapter. The interaction between memory supersession and adapter updates is, to our knowledge, not systematically studied.
OPPU already combines a user adapter with retrieval and profile augmentation, and SPRInG combines a continually updated adapter with gated retrieval. These show that hybrids are feasible and often outperform single-source variants on their own benchmarks. Open design questions are division of labor (which knowledge belongs in memory and which in weights), arbitration at inference (prompt-level conditioning versus logit-level interpolation), and consistency (what happens when memory and weights disagree).
A plausible division, offered here as a hypothesis and not an established result: facts that must be exact, editable, or deletable (names, dates, identifiers) stay in memory, while diffuse regularities (tone, formatting, recurring workflow patterns) are candidates for weights.
| Family | Examples | What it measures | Limits for continual personalization |
|---|---|---|---|
| Personalized generation and classification | LaMP (Salemi et al., 2024), LongLaMP† | Output quality given a user's history | Documents written by users, not assistant dialogue; limited feedback signal |
| Long-term conversational memory | LongMemEval (Wu et al., 2025), LoCoMo† | Recall, temporal reasoning, knowledge updates | Evaluates memory systems, not parametric adaptation |
| Simulated preference users | PersonalLLM† | Adaptation under sparse feedback | Synthetic users |
| Continual-learning protocols | Surveyed in Shi et al. (2024), Yang et al. (2024) | Forgetting, transfer | Not designed around individual users |
Metrics to consider. Preference accuracy, correction recurrence (does a corrected mistake return), consistency over time, temporal and knowledge-update accuracy, retention of general capability on standard benchmarks, adapter-induced behavioral drift, training cost, latency, and storage. Most of these are drawn from the benchmark families above, so we do not claim novelty for individual metrics.
Observation. Zhang et al. (2025) call for new benchmarks and metrics designed for personalized outputs, and Liu et al. (2025) call for benchmarks that include lifelong updates. A benchmark that combines conversational feedback streams, preference change, and measured retention of general capability would address both calls.
Local execution. Keeping data and adapters on the user's device avoids transmitting personal data to a provider. OPPU's authors cite model ownership and privacy as a motivation for personal adapters (Tan et al., 2024). Whether adapters trained on personal data leak that data is a separate question: one 2025 preprint notes in its ethics statement that it did not evaluate membership-inference risk, which illustrates how rarely this is tested. [Search for extraction and membership-inference studies on LoRA adapters.]
Compute. Quantized low-rank training makes single-user fine-tuning of 7B to 8B models plausible on a single consumer GPU with roughly 16 GB of memory, though the exact fit depends on sequence length, batch size, and framework, and should be measured and not assumed. Repeated updates multiply this cost, and latency before new information takes effect is a real disadvantage compared with memory.
Other risks. Reinforcing user errors or biases, over-reliance on a model that mirrors the user (PersonalLLM† discusses feedback loops and personification), difficulty deleting learned content, and drift away from safe or truthful behavior as the adapter accumulates updates.
The table states design-level properties and reasoned expectations, not measured results.
| Dimension | A Base | B Memory/RAG | C Parametric | D Hybrid |
|---|---|---|---|---|
| User knowledge stored in | Nowhere | External store | Adapter weights | Both |
| Parameters change | No | No | Yes | Yes |
| Time to apply new information | N/A | Immediate | After a training step | Immediate via memory, later via weights |
| Editing or deleting one fact | N/A | Easy | Hard | Easy in memory, hard in weights |
| Capture of implicit style | None | Limited | Plausibly stronger | Plausibly stronger |
| Forgetting or drift risk | None | None | Present | Present |
| Storage growth | None | With history | Fixed per adapter | Both |
| Inference overhead | None | Retrieval and longer prompts | Adapter loading | Both |
| Privacy exposure | Low | Content readable if store is compromised | Content may be memorized in weights | Both |
What has been compared experimentally. Within the sources reviewed, OPPU compares per-user adapters against prompt-based retrieval and profile baselines on LaMP and reports combinations with them, and SPRInG compares a continual hybrid against standard RAG and continual-learning baselines on long-form LongLaMP tasks. We did not find a study that evaluates all four architectures together on conversational feedback streams with long-term memory evaluation, general-capability retention, and cost measured in the same protocol. This statement is limited to sources reviewed so far and must be re-checked after the full search.
Competing claims. Parametric methods are reported to outperform prompt-based ones on LaMP (Tan et al., 2024), yet parametric updates are described as expensive and forgetting-prone, and memory as weak at scale (Kim and Kim, 2026). These are not contradictory, but they are measured on different tasks, so they do not establish when each wins.
| Candidate gap | Existing work | Status | Judgment |
|---|---|---|---|
| Distinguishing stable preferences from transient instructions | Novelty scoring in SPRInG; memory write/supersede logic | No dedicated treatment found; confirm | Promising and feasible with a labeled conversational dataset |
| Quality-aware selection of training interactions | SPRInG's drift-driven selection | Partly addressed, on document tasks | Extension to feedback-bearing dialogue is open |
| Feedback-driven continual adapter training | SPRInG, COPE† | Active; first results exist | Differentiation needed; evaluation setting is the opening |
| Memory plus parametric hybrids | OPPU, SPRInG | Established in principle | Open: division of labor, conflict resolution |
| Contradiction and preference change across memory and weights | Memory supersession (LongMemEval); drift in SPRInG | Not systematically studied for weights | Promising, needs a new test design |
| Forgetting during repeated per-user updates | CL surveys; O-LoRA, EWC-LoRA as baselines in SPRInG | Methods exist; user-level evaluation thin | Measurable and feasible |
| Long-horizon evaluation | LongMemEval (memory only), LaMP time splits | Partial | High value; benchmark creation is substantial work |
| Local privacy-preserving pipelines | OPPU motivation; consumer-GPU QLoRA | Empirical privacy analysis thin | Feasible case study |
Priority directions. (1) A benchmark with simulated and real user streams containing durable preferences, transient instructions, and preference reversals, scored for memory, weights, and hybrids. (2) A study of adapter forgetting and behavioral drift over many sequential updates, with general-capability retention measured at each step. (3) Controlled comparisons of A, B, C, and D under identical data, hardware, and budget. (4) Privacy audits of personal adapters.
This section describes a design under development and is not an evaluated contribution. The author's prototype (PAI) runs Qwen3 8B locally through Ollama, with SQLite long-term memory (fact, preference, goal, habit, and project types; add, update, ignore, and supersede operations), logging of interactions with retrieved context and outcomes, detection of explicit corrections and feedback, and a deterministic score estimating each interaction's value as personalization evidence. Planned stages are: build user-specific training sets from high-scoring interactions, fine-tune a LoRA/QLoRA adapter, evaluate it against the memory-only system, keep learning from new interactions, and monitor forgetting and drift.
In terms of the taxonomy, this is a hybrid design (memory plus adapter), with update signal from explicit feedback, and a periodic batch update regime that could move toward continual updates. Its closest published relatives are OPPU and SPRInG, so its value as research lies not in the architecture itself but in a conversational feedback setting, a quality filter tied to preference durability, and a controlled A/B/C/D evaluation on consumer hardware.
Personalizing LLMs has moved from prompt-level conditioning toward parameter-level and hybrid approaches, and from static adaptation toward continual adaptation. Memory and retrieval give editable, immediate personalization. Adapters can capture diffuse patterns and give user ownership, at the price of training cost, deletion difficulty, and forgetting. Early hybrid and continual methods suggest the two are complementary. The main open problems are evaluation on long conversational feedback streams, distinguishing durable from transient preferences, controlling forgetting and drift across repeated updates, and privacy analysis of personal adapters. Closing them requires controlled comparisons that the field does not yet have.
Entries without † were checked against a primary or publisher page during drafting. Entries with † still need verification of authors, venue, and year.
Kim, S., and Kim, J. (2026). SPRInG: Continual LLM Personalization via Selective Parametric Adaptation and Retrieval-Interpolated Generation. arXiv:2601.09974. Preprint.
Liu, J., Qiu, Z., Li, Z., Dai, Q., Zhu, J., Hu, M., Yang, M., and King, I. (2025). A Survey of Personalized Large Language Models: Progress and Future Directions. arXiv:2502.11528. Preprint.
Salemi, A., Mysore, S., Bendersky, M., and Zamani, H. (2024). LaMP: When Large Language Models Meet Personalization. Proceedings of the 62nd Annual Meeting of the ACL (Volume 1: Long Papers), pp. 7370–7392. doi:10.18653/v1/2024.acl-long.399.
Shi, H., Xu, Z., Wang, H., Qin, W., Wang, W., Wang, Y., Wang, Z., Ebrahimi, S., and Wang, H. (2024). Continual Learning of Large Language Models: A Comprehensive Survey. arXiv:2404.16789.
Tan, Z., Zeng, Q., Tian, Y., Liu, Z., Yin, B., and Jiang, M. (2024). Democratizing Large Language Models via Personalized Parameter-Efficient Fine-tuning. EMNLP 2024 (Main). arXiv:2402.04401.
Westhäußer, R., et al. (2025). Enabling Personalized Long-term Interactions in LLM-based Agents through Persistent Memory and User Profiles. arXiv:2510.07925. Preprint.
Wu, D., Wang, H., Yu, W., Zhang, Y., Chang, K.-W., and Yu, D. (2025). LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory. ICLR 2025. arXiv:2410.10813.
Yang, Y., Zhou, J., Ding, X., Huai, T., Liu, S., Chen, Q., Xie, Y., and He, L. (2024). Recent Advances of Foundation Language Models-based Continual Learning: A Survey. ACM Computing Surveys. doi:10.1145/3705725.
Zhang, Z., Rossi, R. A., Kveton, B., Shao, Y., Yang, D., Zamani, H., Dernoncourt, F., et al. (2025). Personalization of Large Language Models: A Survey. Transactions on Machine Learning Research. arXiv:2411.00027.
† Dettmers, T., et al. (2023). QLoRA: Efficient Finetuning of Quantized LLMs. NeurIPS 2023.
† Hu, E. J., et al. (2022). LoRA: Low-Rank Adaptation of Large Language Models. ICLR 2022. arXiv:2106.09685.
† Kirkpatrick, J., et al. (2017). Overcoming catastrophic forgetting in neural networks. PNAS.
† Kumar, I., et al. (2024). LongLaMP benchmark for personalized long-form text generation.
† Maharana, A., et al. (2024). Evaluating Very Long-Term Conversational Memory of LLM Agents (LoCoMo). arXiv:2402.17753.
† PersonalLLM: Tailoring LLMs to Individual Preferences. arXiv:2409.20296.
† PEFT-U: Parameter-Efficient Fine-Tuning for User Personalization.
† Knowledge Graph Tuning: Real-time Large Language Model Personalization based on Human Feedback.
† COPE: Continual Personalization of LLMs under Sparse User Feedback via User Embeddings and Self-Evaluation.
† Beyond Retrieval: Learning Compact User Representations for Scalable LLM Personalization (TAP-PER). arXiv:2606.04547.
† Zhang et al. (2025), PROPER: hierarchical user grouping with MoE-LoRA.