---
title: Epistemic Forgetting
url: https://www.emergentmind.com/topics/epistemic-forgetting
type: topic
---

# Epistemic Forgetting

Epistemic forgetting denotes processes through which a knowledge-bearing system loses usable access to information, weakens the force of prior evidence, or ceases to preserve the distinctions that made a proposition, variable, skill, or reasoning framework action-guiding. Recent work treats it not as a single operation but as a heterogeneous family: loss of retrievability and discriminability in semantic spaces, discounted evidential force in sequential inference, downskilling in logical models, collapse of framework-selection boundaries in modular agents, and state-transforming operations such as contraction, ignoration, revocation, marginalization, and conditionalization on epistemic states [2604.06222] [2601.09726] [2410.22763] [2603.14799] [2508.21441]. This suggests that epistemic forgetting is best understood as a structural transformation of what can be known, retrieved, inferred, or responsibly acted upon, rather than as mere deletion.

## 1. Conceptual scope and principal distinctions

A central recent clarification treats forgetting as “an umbrella term for sociotechnical processes through which something becomes unavailable, varying in agency (intentional vs. incidental), temporality (momentary vs. durable), reversibility (recoverable vs. irreversible), and scale (individual records vs. population-level).” Within that umbrella, **erasure** is “an operational act that removes or disables access to data artifacts,” **unlearning** is “an intentional technical intervention aimed at removing (or bounding) the influence of specific data on learned parameters and downstream outputs,” and **exclusion** is “upstream decisions not to collect, label, or include certain people, events, or categories in datasets” [2602.21180]. In parallel, a general framework of epistemic forgetting on epistemic states distinguishes five abstract types: **contraction** (\(\Psi^\circ_A \not\models A\)), **ignoration** (\(\Psi^\circ_A \not\models A\) and \(\Psi^\circ_A \not\models \overline A\)), **revocation** (\(\Psi^\circ_A \models \overline A\)), **marginalization** (\(\Psi^\circ_A=\Psi_{\Sigma\setminus \Sigma_A}\)), and **conditionalization** (\(\Psi^\circ_A=\Psi\mid \overline A\)) [2508.21441].

Machine-learning work introduces a further operational distinction between **selective forgetting**, treated as beneficial and intentional, and **catastrophic forgetting**, treated as harmful loss of previously learned information. The survey literature organizes selective forgetting by **content of forgetting** (item-, feature-, class-, task-, and stream-level), **recoverability of forgetting** (irrecoverable versus recoverable, including ephemeral, inhibitory, and gradient forgetting), and **extent of forgetting** (exact versus approximate) [2405.20620]. Exact forgetting is defined counterfactually: the post-forgetting model should behave like a model never trained on the removed data. Approximate forgetting weakens that requirement to practical attenuation of influence.

At the representational level, some work defines epistemic forgetting not as deletion of stored content but as loss of access. In "The Geometry of Forgetting" [2604.06222], forgetting is explicitly framed as loss of *retrievability* and *discriminability* within a meaning-structured space: what is “forgotten” is often not erased from storage, but becomes inaccessible, outcompeted, or confusable with nearby alternatives. This yields a broad conceptual convergence: epistemic forgetting often concerns the transformation of availability, not simply the destruction of traces.

## 2. Retrieval geometry, evidence decay, and uncertainty management

One influential line models forgetting as a property of retrieval geometry. In the geometric account, each memory item is stored as an embedding vector produced by a frozen pre-trained encoder, optionally concatenated with contextual information and projected through a trained linear layer with LayerNorm. Retrieval is similarity-based, with score
$$
\text{score} = \cos(\mathbf{q}, \mathbf{m}) \times (1 + \beta t)^{-\psi},
$$
with \(\psi = 0.5\), and forgetting is measured as decline in retrieval accuracy as a function of age and competitor load [2604.06222]. The paper’s core comparative result is that temporal decay alone and no competitors yields a fitted forgetting exponent \(b \approx 0.009\), whereas adding \(10{,}000\) distractors with the same temporal decay function yields \(b = 0.460 \pm 0.183\), with 95% CI \([0.354, 0.644]\) and \(R^2 = 0.757 \pm 0.058\), close to the human benchmark \(b \approx 0.5\). The same work reports that production embedding models with nominal dimensions \(384\), \(768\), and \(1{,}024\) have participation-ratio effective dimensionalities \(15.7 \pm 0.0\), \(16.6 \pm 0.1\), and \(16.3 \pm 0.1\), respectively, implying that interference is governed by effective rather than advertised dimensionality. False memory appears as the complementary effect: with BGE-large on the 24 standard DRM lists, the critical-lure false alarm rate is \(0.583\) at threshold \(\theta=0.82\), compared with a human rate of approximately \(0.55\), and the threshold was chosen because it yields zero unrelated false alarms [2604.06222]. In this picture, forgetting and distortion are coupled consequences of proximity-based retrieval in crowded semantic neighborhoods.

A second line treats forgetting as discounted evidence integration. "Forgetting as a Feature: Cognitive Alignment of Large Language Models" formalizes epistemic forgetting as discounted Bayesian updating,
$$
p_t(\theta \mid \mathcal{D}_{1:t}) \propto p(D_t \mid \theta)\,[p_{t-1}(\theta \mid \mathcal{D}_{1:t-1})]^\gamma,
$$
with \(\gamma \in (0,1]\), and operationalizes prompting-time forgetting by exponential recency weights
$$
w_i = \frac{\exp(-\lambda (t-i))}{\sum_{j=1}^t \exp(-\lambda (t-j))}.
$$
The stated interpretation is that forgetting is not discrete deletion but attenuation of prior evidential force, supporting a stability–plasticity trade-off under nonstationarity and finite resources [2601.09726]. The same paper introduces benchmark families—Temporal Recall, Concept Drift, and Associative Memory—but also notes, in effect, that the forgetting-specific empirical case is limited by mismatches in the reported experiments.

A third line uses explicit uncertainty-carrying belief states. In "The Silent Scholar Problem" [2512.20884], each proposition \(p_i\) is tracked by a Beta posterior with discounted pseudo-counts
$$
\alpha_t=\gamma\alpha_{t-1}+y_t,\qquad
\beta_t=\gamma\beta_{t-1}+(1-y_t),
$$
so that
$$
\mathbb{E}[\theta]=\frac{\alpha}{\alpha+\beta}, \qquad
\mathrm{Var}(\theta)=\frac{\alpha\beta}{(\alpha+\beta)^2(\alpha+\beta+1)}.
$$
The effective sample size satisfies
$$
N_{\text{eff},t}=\gamma N_{\text{eff},t-1}+1,
$$
with equilibrium
$$
N_{\text{eq}}=\frac{1}{1-\gamma}.
$$
The paper argues that this prevents asymptotic collapse of epistemic uncertainty and creates a “homeostatic motive” for external interaction: old evidence loses force, uncertainty re-emerges, and agents have a standing incentive to seek fresh feedback [2512.20884]. A plausible implication is that epistemic forgetting can be engineered not only as loss but as a control mechanism for continual belief maintenance.

## 3. Discriminatory capacity, downskilling, and framework selection

Another major tradition models epistemic forgetting as loss of discriminatory or inferential competence rather than weakening of stored content. "Epistemic Skills: Logical Dynamics of Knowing and Forgetting" introduces weighted models \(M=(W,E,C,B)\), where \(E(w,u)\) is a skill set attached to each world-pair and \(C(a)\) is the capability set of agent \(a\). Individual knowledge is defined by
$$
M,w \models K_a \varphi \iff \forall u\in W \bigl(C(a)\subseteq E(w,u)\Rightarrow M,u\models \varphi\bigr),
$$
so skills determine which worlds an agent can distinguish. Forgetting is then formalized as **downskilling**:
$$
M,w \models (-S)_a \varphi \iff (W,E,C_{a-S},B),w \models \varphi,
$$
where \(C_{a-S}(a)=C(a)\setminus S\) [2410.22763]. The underlying ontology is explicit: forgetting is not deletion of facts or alteration of valuation \(B\), but reduction of discriminatory capacity. Knowledge can be lost because more worlds become epistemically accessible once skills are removed. Group modalities \(E_G\), \(C_G\), \(D_G\), and \(F_G\) extend this to mutual, common, distributed, and field knowledge, so forgetting can propagate into social epistemic structure without being primitive group-forgetting.

A closely related but architecturally distinct notion appears in "Universe Routing: Why Self-Evolving Agents Need Epistemic Control" [2603.14799]. There, a “belief space universe” is defined as a tuple \(u=(A_u,I_u,S_u)\) of axioms, inference procedure, and solver, and universes are epistemically incompatible when one contains assumptions equivalent to the negation of the other. The router solves
$$
u^* = \arg\max_{u \in \mathcal{U}} P(u \mid q),
$$
over \(K=7\) universes, including `STAT_FREQ`, `STAT_BAYES`, `PHYS_CLASSICAL`, `PHYS_QUANTUM`, `PHYS_RELATIVITY`, `STAT_MIXED`, and `STAT_ILL_POSED`. In this setting, epistemic forgetting is not loss of factual content but collapse of previously learned universe boundaries. The consequence is “wrong in kind”: the system invokes the wrong epistemology rather than merely producing a noisier answer. Empirically, hard routing and soft routing both achieve \(97.25\%\) accuracy, but hard routing takes \(5.5\)ms versus \(38.2\)ms; in continual learning from 5 to 7 universes, Naive Fine-tuning yields \(11.84\%\) old-universe accuracy and \(86.84\%\) forgetting, EWC yields \(23.68\%\) and \(75.00\%\), while Rehearsal (10\%) reaches \(98.68\%\) old-universe accuracy and \(0.00\%\) forgetting [2603.14799]. The paper’s interpretation is that epistemic forgetting here is forgetting *how to reason*—specifically, forgetting which incompatible reasoning framework a query belongs to.

Taken together, these models shift the object of forgetting from propositions to capacities. This suggests that epistemic forgetting may consist in losing the resources that keep alternatives apart: skill sets in logical models, or decision boundaries between incompatible inferential spaces in modular agents.

## 4. Formal operators on epistemic states, Kripke models, and compiled languages

A substantial formal literature studies epistemic forgetting as an operation on explicit epistemic structures. In knowledge-structure semantics, a model is
$$
\mathcal{F}=(V,\Gamma,O_1,\dots,O_n),
$$
with background theory \(\Gamma\) and observable variables \(O_i\) for each agent. Agent \(i\)’s knowledge is truth across all states agreeing on \(O_i\), and forgetting appears as elimination of unobservable variables. The key result is that, when \(\Gamma=\{\theta\}\),
$$
(\mathcal F,s)\models K_i\alpha \iff (\mathcal F,s)\models \forall(V\setminus O_i)(\theta\rightarrow \alpha),
$$
so knowledge is characterized by the weakest sufficient condition of \(\alpha\) over \(O_i\); common knowledge is characterized by a generalized weakest sufficient condition expressed by a greatest fixed point [1401.3472]. Here forgetting is not a dynamic event but the computational core of observability-based epistemic reasoning.

At the level of epistemic states with richer semantics, "Forgetting Formulas and Signature Elements in Epistemic States" shows that marginalisation is the correct extension of Delgrande-style forgetting of signature elements to ordinal conditional functions. For \(\Sigma'\subseteq \Sigma\),
$$
\kappa_{|\Sigma'}(\omega')=\min\{\kappa(\omega)\mid \omega\in\Omega_\Sigma,\ \omega\models \omega'\},
$$
and the induced beliefs satisfy
$$
Bel(\kappa_{|\Sigma'}) = Bel(\kappa)\cap \mathcal{L}_{\Sigma'}.
$$
The same paper proves a sharp negative result: extending analogous postulates to forgetting arbitrary formulas in epistemic states forces triviality, so that \(Bel(\Psi L \varphi)\equiv \top\) for each \(\varphi\) [2108.08603]. In this sense, variable/signature forgetting and formula forgetting have fundamentally different rationality profiles.

Dynamic-epistemic treatments of forgetting whether a complex proposition use more interventionist model transformations. "Forgetting complex propositions" defines a modality
$$
[\boldsymbol{\ddagger}\,\pi]\varphi
$$
for forgetting whether \(\pi\), based on the prime implicate clausal forms \(C(\pi)\) and \(C(\neg\pi)\). A model transformation \(M_u^{\mathcal C}\) creates clause-falsifying copies of the original model, and
$$
M,w \models [\boldsymbol{\ddagger}\,\pi]\varphi
$$
holds iff \(\varphi\) holds after every minimal way of falsifying one clause in \(C(\pi)\) and one clause in \(C(\neg\pi)\) [1507.01111]. Because different minimal choices can yield non-bisimilar outcomes, forgetting whether a complex formula is inherently non-unique. The same paper represents the construction in action model logic and gives a sound and complete axiomatization.

Compiled representations provide a different route to tractable forgetting. In multi-agent epistemic logic, forgetting is defined as variable elimination:
$$
\exists Q.\phi,
$$
the strongest consequence of \(\phi\) over the reduced variable set \(P\setminus Q\). The key structural condition is **logical separability**, which enables modular forgetting on separable terms:
$$
\exists Q. \phi \equiv \bigwedge_{\alpha \in Prop(\phi)} (\exists Q. \alpha) \land \bigwedge_{i \in B} \left[\bigwedge_{\beta \in \BoxSub[i](\phi)} \Box_i (\exists Q. \beta) \land \bigwedge_{\gamma \in \DiamSub[i](\phi)} \Diamond_i (\exists Q. \gamma)\right].
$$
The compiled normal form SDNF supports tractable forgetting provided the base propositional language supports \(FO\) or \(SFO\) [1806.10561]. This is an algorithmic rather than ontological contribution: epistemic forgetting becomes feasible because epistemic formulas are compiled into a separability-preserving language.

Belief-contraction work in dynamic epistemic logic models forgetting-like uncertainty increase directly on Kripke models. For contraction on \(\varphi\),
$$
R_a^{\div\varphi}(w)=
\begin{cases}
R_a(w)\cup \{v\in W: M,v\vDash \neg\varphi\} & \text{if } M,w\vDash B_a\varphi,\\
R_a(w) & \text{otherwise.}
\end{cases}
$$
The dynamic clause is
$$
M,w \vDash [\div\varphi]\psi \iff M^{\div\varphi},w \vDash \psi.
$$
This models belief loss by re-opening \(\neg\varphi\)-possibilities, and generalized event models add a second relation \(Q_a^+\) to represent newly introduced event possibilities [2606.31861]. The operation is not variable forgetting or forgetting whether, but it is a direct formalization of epistemic loosening.

## 5. Continual learning, modular memory, and machine-learning implementations

In neural continual learning, epistemic forgetting often appears as controlled release of certainty rather than overwrite. "Bayesian continual learning and forgetting in neural networks" introduces MESU, which replaces full sequential Bayes with a truncated posterior over only the last \(N\) tasks and derives a learning–forgetting factorization in which forgetting appears as an inverse-likelihood term for the oldest retained data [2504.13569]. The approximate posterior is mean-field Gaussian,
$$
q_{\bm{\theta}_t}(\bm{\omega}) = \prod_i \mathcal N(\omega_i;\mu_{t,i},\sigma_{t,i}^2),
$$
so forgetting acts by pulling means and variances back toward the prior, preventing vanishing uncertainty and “catastrophic remembering.” On 200 sequential permuted MNIST tasks, MESU slightly outperforms EWC Online and SI on tasks inside its memory window—\(91.37\%\) versus \(88.5\%\) and \(87.0\%\)—while FOO-VB Diagonal shows catastrophic remembering, with new-task accuracy saturating around \(70\%\). In a long-training experiment on MNIST, MESU maintains nonzero mean \(\sigma\) and ROC AUC for OOD detection stays near \(1.0\) [2504.13569]. The paper’s conceptual claim is that epistemic forgetting here is uncertainty-mediated selective de-consolidation.

The broader machine-learning survey literature systematizes this design space. Selective forgetting is divided into **active forgetting**, intended to improve performance, and **passive forgetting**, driven mainly by privacy, security, or usability. Exact forgetting requires that “the model state replicate the state it would have in the complete absence of the forgotten data,” whereas approximate forgetting weakens that requirement to practical attenuation. The same survey treats forgetting as support for generalization, overfitting mitigation, compression, and privacy preservation, and notes that “learning and forgetting should be considered different aspects of a cognitive process of belief updating” [2405.20620]. In attention and long-context models, forgetting may take the form of bounded retention, as in Expire-span, where each memory \(h_i\) is assigned an expire-span \(e_i\) determining how long it remains accessible.

These implementations differ sharply in mechanism—uncertainty release, replay-preserved modularity, recency weighting, memory expiration—but converge on a common principle: epistemic forgetting is often engineered to regulate interference, preserve plasticity, or remove obsolete influence without requiring complete erasure of representational content.

## 6. Governance, unlearning, and epistemic justice

The governance literature insists that epistemic forgetting cannot be reduced to a single technical primitive. Data systems are described as “infrastructures of memory,” and forgetting is said to operate through erasure, unlearning, exclusion, semantic dependencies, and access-limiting arrangements that alter what systems, institutions, and publics can know [2602.21180]. On this view, deleting records from storage does not remove their influence on trained models, and unlearning is therefore only one component of responsible forgetting. The normative proposal is that unlearning become a first-class capability in knowledge infrastructures and be evaluated not only by compliance or utility retention, but also by **transparency**, **accountability**, and **epistemic justice**. The same work emphasizes the dual character of forgetting: it can protect rights and reduce harm, but it can also enable silencing, remove evidence, and reproduce epistemic injustice by erasing underrepresented perspectives.

Survey work on machine forgetting aligns with this normative turn by connecting unlearning to GDPR and CCPA, by treating **completeness** as the requirement that forgotten data leave “no residual trace,” and by emphasizing **data lineage management** as a way to track how evidence enters model knowledge and how its removal should propagate [2405.20620]. Exact versus approximate forgetting is therefore not only a technical distinction but also a governance distinction: exact forgetting approximates counterfactual non-acquisition, while approximate forgetting manages residual influence pragmatically.

A broader implication follows from the contrast between protective and harmful forgetting. Some recent work explicitly argues that forgetting is “not an exception to knowledge, but its condition” [2602.21180]. This suggests that epistemic forgetting is constitutive of knowledge systems: they must decide what becomes unavailable, what remains actionable, and whose traces continue to count as evidence. In computational settings, that decision is simultaneously technical, epistemic, and political.

Source: https://www.emergentmind.com/topics/epistemic-forgetting