When to Forget: A Memory Governance Primitive

Baris Simsek

cs.AIApr 13, 2026

When to Forget: A Memory Governance Primitive

Authors: Baris Simsek

Abstract

Agent memory systems accumulate experience but currently lack a principled operational metric for memory quality governance -- deciding which memories to trust, suppress, or deprecate as the agent's task distribution shifts. Write-time importance scores are static; dynamic management systems use LLM judgment or structural heuristics rather than outcome feedback. This paper proposes Memory Worth (MW): a two-counter per-memory signal that tracks how often a memory co-occurs with successful versus failed outcomes, providing a lightweight, theoretically grounded foundation for staleness detection, retrieval suppression, and deprecation decisions. We prove that MW converges almost surely to the conditional success probability p+(m) = Pr[y_t = +1 | m in M_t] -- the probability of task success given that memory m is retrieved -- under a stationary retrieval regime with a minimum exploration condition. Importantly, p+(m) is an associational quantity, not a causal one: it measures outcome co-occurrence rather than causal contribution. We argue this is still a useful operational signal for memory governance, and we validate it empirically in a controlled synthetic environment where ground-truth utility is known: after 10,000 episodes, the Spearman rank-correlation between Memory Worth and true utilities reaches rho = 0.89 +/- 0.02 across 20 independent seeds, compared to rho = 0.00 for systems that never update their assessments. A retrieval-realistic micro-experiment with real text and neural embedding retrieval (all-MiniLM-L6-v2) further shows stale memories crossing the low-value threshold (MW = 0.17) while specialist memories remain high-value (MW = 0.77) across 3,000 episodes. The estimator requires only two scalar counters per memory unit and can be added to architectures that already log retrievals and episode outcomes.

Figures & tables

Figure 1: Memory Worth calibration over episodes. Spearman $\rho$ between $\mathrm{MW}_{T}(m)$ and $U^{*}(m)$ for all four weighting strategies, averaged over 20 seeds (shaded regions: $\pm$ 1 std). The no-feedback baseline stays at $\rho=0$ throughout. All three updating strategies converge to $\rho\approx 0.89$ by episode 10,000.

Method	$\rho$ @2k	$\rho$ @5k	$\rho$ @10k	$\Delta$
No update	$0.00_{\pm.00}$	$0.00_{\pm.00}$	$0.00_{\pm.00}$	$-0.89$
Uniform	$0.66_{\pm.06}$	$0.81_{\pm.03}$	$0.89_{\pm.02}$	$0.00$
Ours (sim-wt.)	$0.66_{\pm.06}$	$0.81_{\pm.04}$	$0.89_{\pm.02}$	$0.00$
Oracle	$0.67_{\pm.06}$	$0.82_{\pm.04}$	$0.90_{\pm.02}$	$+0.00$

Table 1: Memory Worth calibration summary. Values are Spearman

\rho

(mean

\pm

std over 20 seeds).

\Delta

is the difference from Uniform at episode 10,000. Beta-Bernoulli converges to the same ranking as Uniform at long-run evidence levels and is discussed in the text.

Figure 2: Experiment 2: Task-difficulty confound (20 seeds, mean $\pm$ std). Global MW (orange) remains negatively correlated with true utility ( $\rho\approx-0.33$ ) because specialist memories ( $U^{*}=0.85$ ) appear only on low-success hard tasks and are penalised by task difficulty. A weighted-average conditional MW (yellow) fails equally: mixing easy- and hard-task counts preserves the same base-rate confound. When MW is conditioned on the hard-task population only and evaluated over memories with hard-task evidence (green), the ranking signal becomes positive ( $\rho\approx+0.14$ ), showing partial recovery once task-type confounding is removed. The gap to the unconfounded baseline ( $0.89$ ) remains large, motivating within-stratum normalisation as future work.

Figure 3: Experiment 3: Retrieval policy feedback loop. MW-weighted retrieval converges to $\rho\approx 0.896$ – $0.899$ across all exploration floors, matching or marginally exceeding the uniform reference ( $\rho=0.890$ , 20 seeds, mean $\pm$ std). Even with $\varepsilon=0$ (no forced exploration floor), the estimator does not degenerate. The self-correcting feedback loop outweighs the rich-get-richer effect in this regime.

Figure 4: Experiment 4: Per-memory MW trajectories for anchor ( $U^{*}=0.90$ ) and hitchhiker ( $U^{*}=0.05$ ) under varying independence fractions. Dashed lines mark true $U^{*}$ . With 0% independent retrievals, both memories converge to identical MW $\approx 0.49$ , indistinguishable despite a 17 $\times$ difference in true utility. Separation requires at least 30% independent episodes, and full convergence to true utility only with 100% independence.

Figure 5: Experiment 5: MW trajectories in a text-based retrieval agent ( all-MiniLM-L6-v2 embedding retrieval, 3,000 episodes, 20 seeds, mean $\pm$ std). The stale memory (dashed orange) peaks at $\approx 0.97$ in Phase 1, drops sharply after episode 100, crosses $\theta_{L}=0.40$ near episode 300, and ends at $0.17$ —well into the low-value category, with no sign of convergence, warranting deprecation. The specialist (solid green) stabilises at $0.77$ . The hitchhiker (dotted grey, $0.77$ ) and control (dash-dot blue, $0.73$ ) are both elevated and nearly indistinguishable from the specialist—consistent with the co-retrieval confound of Experiment 4: MiniLM routes Python queries to both specialist and hitchhiker due to their semantic similarity.

Memory type	MW @ ep100	MW @ ep500	MW @ ep3000
Stale	${\approx}0.97$	${\approx}0.40$	$0.17$
Specialist	${\approx}0.93$	${\approx}0.81$	$0.77$
Hitchhiker	${\approx}0.91$	${\approx}0.79$	$0.77$
Control	${\approx}0.78$	${\approx}0.77$	$0.73$

Table 2: Experiment 5 MW summary (values from figure, 20 seeds). all-MiniLM-L6-v2 retrieval, 3,000 episodes; stale crosses

\theta_{L}=0.40

near episode 300.

Agent Memory

Agent memory lets AI systems retain, update, and use facts, experiences, and skills across sessions and long-horizon tasks.

See all 183 papers

Jun 11, 2026

Learning What to Remember: A Cognitively Grounded Multi-Factor Value Model for Agentic Memory

Long-running LLM agents accumulate interaction histories far larger than any context window, forcing a standing decision: what to encode deeply, what to forget, and what to retrieve under a fixed memory budget. Production systems answer with semantic similarity or recency -- both mis-specified for the forgetting decision, which is made at consolidation time before the future query is known. We propose a multi-factor memory value function V(m)=\sum_i w_i f_i(m) over seven interpretable factors (emotional intensity, goal relevance, value alignment, self/user relevance, task utility, reliability, and usage history) drawn from cognitive psychology, whose weights are learned from a downstream objective by a gradient-free optimiser, and whose single scalar uniformly controls encoding depth, forget risk, and retrieval rank. We make a methodological point: on LongMemEval, scoring goal relevance against the held-out evaluation question saturates gold-evidence retention at \approx 0.98 -- this measures retrieval, not forgetting. In the realistic blind regime, a learned multi-factor value retains 0.770 \pm 0.011 of gold evidence across 479 usable cases, versus 0.657 for uniform weights, 0.518 for the best single factor, and 0.368 for recency; every paired gap's 95% bootstrap CI is above zero, and a neural network over the same factors ties the linear model. The learned weights are interpretable -- reliability, emotional intensity, and self/user relevance dominate, while query-time goal similarity is correctly down-weighted for the forgetting decision. A controlled synthetic task with planted confounds confirms the learner recovers a separating weighting (1.00 retention) where uniform weighting fails (0.62). The substrate is open-source; all experiments run on a single CPU with no API calls.

Zhibao Chen, Qian Cheng

Huatai Securities, Nanjing, China

LLM Agent Memory Persistent Memory for Language Models LLM Memory Memorization in Generative Models Memory Consolidation Agent Memory Management

Ask PDF

Oct 1, 2026

Causal Memory Policy: Making Memory Utility Identifiable by Intervening on Retrieval

Memory-augmented large language models must decide which memories to retain, and recent systems do so by estimating each memory's effect on task performance. However, these estimates rely entirely on retrieved memories. When a memory is never retrieved, store-level interventions produce identical outcomes, leaving its utility unidentified. This is a retrieval-level positivity violation, invisible to diagnostics that examine only memory operations. We introduce Causal Memory Policy (CMP), a causal framework that restores identification by intervening on retrieval itself, reserving a fixed number of context slots for memories sampled with known propensities. CMP estimates memory utility by self-normalized inverse propensity weighting under a balanced assignment design. We prove the causal factorization of memory utility through retrieval, the unbiasedness and exact variance of the estimator, and the optimal decision rule under irreversible operations. Empirically, identification fails for 54% of required memories on LongMemEval and 67% on LoCoMo, and the failure persists in a deployed memory system. CMP improves discrimination between required and non-required memories from 0.54 to 0.66 AUC. Finally, we show that identified memory utility alone is insufficient for retention decisions: per-query utility reaches 0.78 AUC on the query for which it is estimated, yet no aggregation available to a retention policy predicts a memory's value on unseen queries. Code is available at: https://anonymous.4open.science/r/cmp-release-D0C3/.

Arman Behnam, Binghui Wang

Department of Computer Science Illinois Institute of Technology Chicago, IL

LLM Memory Causal Inference Agent Memory Management

Ask PDF

Oct 19, 2026

Remembering Has a Price: Learning Memory Worth under Resource Constraints

Lifelong agents continually accumulate and reuse experience to improve after deployment, yet in practice they often operate under tight generation, interaction, and memory budgets. Under such stringent joint constraints, we find that existing lifelong-memory methods can become ineffective, with several sophisticated approaches offering little advantage over simple retrieval or even no-memory agents. This reveals a more fundamental problem: resource scarcity makes each memory decision more consequential. When only a small amount of past experience can be retained or reused and opportunities for reasoning and recovery are limited, retrieving a relevant but ineffective memory can consume scarce context and hinder downstream execution. The key challenge is therefore not simply to retrieve relevant experience, but to determine which experiences are actually worth retaining and reusing. To address this challenge, we propose TokenWorth, a lightweight online memory framework that learns query-dependent memory worth from sparse execution feedback and applies it consistently to memory retrieval and bounded memory management. Across three LifelongAgentBench environments and both local and API-based deployments, TokenWorth achieves the highest task success rate while using the fewest output tokens among the evaluated methods. These results suggest that effective lifelong adaptation under resource constraints depends not on remembering more, but on learning what is worth remembering. Our code is publicly available at https://github.com/AIGNLAI/TokenWorth.

Guannan Lai, Zi-Yang Bai, Jun-Peng Jiang +1

School of Artificial Intelligence, Nanjing University · National Key Laboratory for Novel Software Technology, Nanjing University

Memory Management for LLM Agents Lifelong Learning Agents

Ask PDF

Abstract

Agent Memory

Explore similar work

Learning What to Remember: A Cognitively Grounded Multi-Factor Value Model for Agentic Memory

Causal Memory Policy: Making Memory Utility Identifiable by Intervening on Retrieval

Remembering Has a Price: Learning Memory Worth under Resource Constraints