Locate-Forget-Update Framework

• Locate–Forget–Update is a modular framework that structures model adaptation into three phases: Locate, Forget, and Update.
• The framework employs precise identification metrics and adapter-based interventions to minimize collateral knowledge disruption.
• Applications include language model rewriting, incremental knowledge base consolidation, recommender evolution, and privacy-compliant machine unlearning.

The Locate–Forget–Update framework structures knowledge editing, rule management, or model adaptation as a three-phase process: identification of target structures (Locate), controlled removal or suppression (Forget), and precision revision or consolidation (Update). Its implementations span LLM knowledge rewriting (Ngugi, 9 Aug 2025), incremental knowledge base consolidation (Martínez-Plumed et al., 2015), recommender system evolution (Liu et al., 20 Nov 2025), and machine unlearning for privacy compliance (Chen et al., 2023). The following sections detail theoretical underpinnings, practical mechanisms, measurement protocols, safety implications, and application domains.

1. Theoretical Foundations and Rationale

Locate–Forget–Update emerged as a formalization of cognitive and algorithmic strategies for continual model adaptation and knowledge management. In knowledge bases and rule-learning systems, the cycle supports incremental learning, bounding memory usage and mitigating the stability–plasticity dilemma via staged consolidation and selective forgetting (Martínez-Plumed et al., 2015). In neural architectures, the framework isolates the minimum set of model parameters causally responsible for targeted outputs—propagating interventions only through these loci minimizes collateral damage, particularly catastrophic forgetting during updates (Ngugi, 9 Aug 2025, Liu et al., 20 Nov 2025).

In machine unlearning, Locate–Forget–Update executes privacy-compliant knowledge removal by externalizing forget operations into lightweight adapters whose optimization is paired with parameter selection and post-hoc fusion (Chen et al., 2023). In compact LLMs, mechanistic interpretability metrics such as activation magnitude, output patching, and gradient norm support the localization step, ensuring that only circuits critical for factual recall are disrupted (Ngugi, 9 Aug 2025). The general principle is precise intervention with controlled scope, avoiding broad retraining and retaining essential unrelated information.

2. Canonical Workflow: Phases and Algorithms

2.1 Locate

In rule-based systems, “Locate” ranks candidate rules by description length, coverage, support, complexity, optimality, and permanence metrics, often derived from Minimum Message Length (MML) formulations on directed coverage graphs (Martínez-Plumed et al., 2015). In transformers, it entails ranking modules—attention heads, MLP sublayers, or low-rank adapters—by causal contribution to the target output (Ngugi, 9 Aug 2025).

In sequential recommendation frameworks, sensitive layers are identified by aggregating hidden state divergence scores (cosine similarity between old vs. new representations) across users, then thresholding to select layers most responsible for preference drift (Liu et al., 20 Nov 2025). Machine unlearning frameworks instrument adapters at every block and log activations per deletion request to enable downstream fusion (Chen et al., 2023).

2.2 Forget

Rule systems remove low-permanence candidates from the working space (short-term memory) until storage bounds are met, redistributing residual support and preserving hierarchical knowledge relationships (Martínez-Plumed et al., 2015). In LLM knowledge rewriting, selected modules are augmented with per-module scaling adapters (e.g., IA³) that inhibit activations causally linked to the target fact, trained to suppress undesired output while regularizing magnitudes (Ngugi, 9 Aug 2025).

Recommender architectures prune outdated items from history using lightweight, parallel sequential filter models, optimizing for relevance with respect to the user's new preference embedding (Liu et al., 20 Nov 2025). Machine unlearning adapters are optimized via paired teacher–student objectives: KL-divergence selectively accentuates retention and suppression, augmented by reversed pretraining loss and downstream task loss for retained data (Chen et al., 2023).

2.3 Update

Knowledge bases promote rules with optimality beyond a threshold into consolidated knowledge, and demote outdated consolidations for potential future forgetting; thresholds can induce hysteresis for stability (Martínez-Plumed et al., 2015). In compact LLMs, the Update phase injects new facts via freshly trained scaling adapters, merged with prior adapters to produce the final circuit-modulated model (Ngugi, 9 Aug 2025).

EvoRec updates only the sensitive subset of pretrained LoRA adapters, optimizing a joint objective aligned to new-user data and regularized for consistency with inactive-user outputs, enforcing a binary mask on trainable parameters (Liu et al., 20 Nov 2025). In unlearning architectures, multiple adapters are fused post hoc by solving a least-squares matrix equation using the stored Gram and cross-term matrices for each layer, ensuring efficient adaptation after sequential deletion requests (Chen et al., 2023).

3. Quantitative Metrics and Evaluation Protocols

Each implementation features rigorously defined evaluation regimes:

• Rule-system consolidation tracks population and accuracy metrics for both short-term (W) and long-term (K) pools, leveraging coverage graphs in domains such as chess move learning (Martínez-Plumed et al., 2015).
• LLM knowledge editing employs fact accuracy, forget rate, and control-fact preservation metrics:

$$\text{ForgetRate} = 1 - \frac{A_{F1}^{\mathrm{post}}}{A_{F1}^{\mathrm{pre}}}$$

$$A_{F2} = \frac{\#\,\text{“Google.” responses}}{200}$$

$$A_{\mathrm{control}} = \frac{\#\,\text{correct control answers}}{100}$$

The framework achieves 98.5% new-fact accuracy, 96% forget rate, and 72% control-fact accuracy, outperforming direct fine-tuning in retention of control facts (Ngugi, 9 Aug 2025).

• EvoRec benchmarks cover HR@1/3, NDCG@3, update runtime, and LoRA parameter footprint, showing ~30% adapter update coverage and substantial computational savings versus full retraining (Liu et al., 20 Nov 2025).
• Efficient Unlearning evaluates test-set accuracy, retained-set fidelity, forgotten-set suppression, MLM loss on deleted data, and wall-clock update time; it demonstrates superior forget-set performance and minimal retention loss compared to SISA, reverse-gradient, and MEND baselines (Chen et al., 2023).

4. Mechanisms of Catastrophic Forgetting Mitigation

Locate–Forget–Update achieves robust catastrophic forgetting mitigation through localized updates and explicit preservation strategies. In compact LLM knowledge editing, interpretability-guided selection of modules ensures only the circuits encoding the target fact are suppressed, minimizing interference with unrelated knowledge (Ngugi, 9 Aug 2025). IA³ and LoRA PEFT schemes operate only on adapter parameters, preserving backbone semantics and reverting to default factual configurations when necessary.

Recommender systems utilize KL-consistency regularization to maintain performance for inactive users, preventing drift during active-user preference adaptation (Liu et al., 20 Nov 2025). Incremental consolidation cycles in rule-based systems isolate high-value generalizations for long-term retention, addressing plasticity without objectionable forgetting (Martínez-Plumed et al., 2015). Machine unlearning architectures operate external to the frozen backbone, leveraging fusion of adapter parameters to continually accommodate deletion requests without accumulation of inference overhead or semantic degradation (Chen et al., 2023).

5. Safety, Reversibility, and Soft Forgetting

Locate–Forget–Update enables nuanced forms of “soft forgetting.” In compact LLMs, knowledge suppression via adapters renders facts conditionally accessible—default retrieval is suppressed, but latent circuits remain reactivatable under explicit cues or safe, context-rich prompting (Ngugi, 9 Aug 2025). This supports reversibility, auditability, and minimizes risk of residual misinformation leakage under adversarial prompting.

In privacy-oriented unlearning, the post-fusion adapters yield compliance without semantic collapse. Retention steps maintain overall model performance, and negative pretraining loss prevents easy recovery of deleted content (Chen et al., 2023). In rule frameworks, demotion cycles allow historical rules to re-enter plastic memory, further supporting recovery and reuse in evolving domains (Martínez-Plumed et al., 2015).

6. Application Domains and Case Studies

Locate–Forget–Update is demonstrated across several modalities:

• Incremental relational rule learning (chess moves): the framework ensures rapid convergence to minimal generalizing rule sets, supports transfer across subtasks, and bounds memory footprint (Martínez-Plumed et al., 2015).
• Surgical factual edits in compact transformers: extensive mechanistic evaluations show controlled update dynamics, high new-fact accuracy, and preservation of knowledge scope beyond the targeted fact (Ngugi, 9 Aug 2025).
• Large-scale sequential recommendation: EvoRec models user drift efficiently, enabling rapid fine-tuning for active users while guarding inactive user interests with only partial adapter updates (Liu et al., 20 Nov 2025).
• Machine unlearning for legal compliance: efficient unlearning adapters allow localized, scalable data deletion, maintaining test performance and procedurally integrating sequential unlearning via parameter fusion (Chen et al., 2023).

7. Limitations and Open Directions

Key challenges for Locate–Forget–Update include refinement of localization metrics—e.g., extending beyond activation and divergence scores to incorporate gradient or Fisher saliency measures—and formal quantification of soft forgetting’s boundary conditions and adversarial resilience. EvoRec suggests future extensions with adaptive parameter selection per user and cross-task localization for group-level drift (Liu et al., 20 Nov 2025). Recurrent fusions and demotion/promotion cycles in rule learning warrant formal study for continuous domain adaptation. A plausible implication is that further advances in mechanistic interpretability and modular fine-tuning will extend Locate–Forget–Update applicability into more granular, multi-fact, and multi-task knowledge management regimes.

---

Locate–Forget–Update thus constitutes a modular, interpretable paradigm for precise, efficient, and safe model adaptation and knowledge management across symbolic, neural, and hybrid systems. Its principled division into localization, selective forgetting, and update phases supports stability–plasticity balance, mitigates catastrophic loss, and accommodates legal and practical constraints across domains (Martínez-Plumed et al., 2015, Ngugi, 9 Aug 2025, Liu et al., 20 Nov 2025, Chen et al., 2023).