LifeFuse-Mem: Lifecycle-Aware State Fusion Against Temporary Overwriting for Long-Term Memory

1. Introduction

• LifeFuse-Mem separates information lifecycles at write time and uses phase-aware readout to preserve long-term facts while following local assumptions.The paper frames its evaluation as a controlled setting with lifecycle metadata and known phase structure.Source: Figure 1 · p. 2, §1 · p. 2
• When transient assumptions and persistent knowledge share a memory pathway, temporary updates can overwrite durable knowledge and silently alter future behavior.Such failures can degrade behavioral consistency and undermine user trust.Source: §1 · p. 2
• LifeFuse-Mem improves acquisition-controlled retention on the direct anti-overwrite benchmark and remains broadly competitive on LoCoMo and MemoryAgentBench.The public benchmarks are compatibility checks; they lack overwrite phases and query lifecycle labels.Source: §1 · p. 3

2.1. Memory-Augmented Language Models

• LifeFuse-Mem is closest to parametric and adapter memories, extending compact online associative states with lifecycle-aware subspace isolation and phase-aware readout.Other memory approaches include retrieval, history compression, and tiered memory management.Source: §2.1 · pp. 3–4

2.2. Continual Learning and Catastrophic Forgetting

• Unlike continual-learning methods that retain prior task behavior, LifeFuse-Mem distinguishes locally valid temporary information from permanent replacement within one online memory state.Continual-learning approaches include replay, regularization, parameter isolation, and subspace or gradient projection.Source: §2.2 · p. 4

2.3. Long-Term Memory Evaluation

• General long-term memory benchmarks typically do not separate permanent facts from temporary overwrites or condition retention on successful initial acquisition.Hard Attribution Anti-Overwrite complements them with Phase-A/Phase-B conflicts and acquisition-controlled overwrite metrics.Source: §2.3 · p. 4

3.1. Memory-Conditioned Response Generation

• Permanent facts encode stable user facts or durable preferences, whereas temporary facts encode local assumptions, task-specific conditions, or counterfactual instructions for the current context.The response-generation formulation combines the parametric model with accumulated memory.Source: §3.1 · p. 4
• A permanent query should recover its durable value after later local context is written, while a temporary query should follow the active temporary value.The hard case is a shared semantic attribute assigned conflicting permanent and temporary values.Source: §3.1 · p. 4

3.2. Lifecycle Conflict in Compact Online Memory

• The shared compact online state uses a token-level write gate and keep coefficient in a delta-style update, an abstraction that includes δ-Mem.The state is injected back into model computation.Source: §3.2 · p. 5
• Temporary overwrite occurs when a temporary write makes a later permanent query return the temporary value instead of the permanent value.Using the same rows and write dynamics across phases can erase or re-rank evidence for the earlier value.Source: §3.2 · p. 5
• Retention is measured only on permanent facts answered correctly after Phase A, separating loss after temporary writes from failure to acquire the fact initially.The protocol defines A_i as correct after Phase A and R_i as returning the permanent value after Phase B.Source: §3.2 · p. 5
• The overwrite rate counts answers matching the conflicting temporary value on the same acquired subset.This tests whether a temporary write becomes authoritative after a permanent fact has already been acquired.Source: §3.2 · p. 5

4.1. Framework Overview

• The framework combines a learned plastic/stable partition, supervised permanent routing, and phase-aware protected-state readout to evaluate recovery after conflicting temporary writes.Protected-state readout is specific to the controlled overwrite-attribution protocol; it is not external key–value lookup.Source: §4.1 · p. 5

4.2. Lifecycle-Aware Subspace Partition

• LifeFuse-Mem learns an approximately orthogonal basis, rotates memory vectors into it, and divides its coordinates into plastic and stable subspaces.The experiments use rank 8, split evenly into four plastic and four stable coordinates; the regularizer keeps the subspaces separable while allowing a learned coordinate system.Source: §4.2 · pp. 5–6

4.3. Permanent Routing and Subspace Writing

• Permanent-route weights scale writes to stable rows, while plastic rows remain writable for every token.In the main runs, temporary writes primarily update plastic rows; stable-key coordinates are also attenuated during ordinary online writing.Source: §4.3 · p. 6

4.4. Episodic Learning Objective

• Episodic training disables further writing during the answer, with cross-entropy, routing, orthogonality, candidate-ranking, retention-ranking, and sparsity terms in the objective.The hard-candidate loss is ramped in after early answer learning stabilizes; the retention-ranking loss pressures permanent answers to outrank temporary or rejected candidates.Source: §4.4 · p. 7

4.5. Protected State Fusion and Lifecycle Readout

• In the controlled protocol, the protected state combines post-overwrite plastic rows with pre-overwrite stable rows, preserving intended long-lived evidence alongside recent context.For permanent queries, candidates are scored under both the pre-overwrite and protected states; temporary queries use the full post-overwrite state without protection or fusion.Source: §4.5 · pp. 7–8

5.1. Experimental Setup

• Comparisons use matched rank-8 LoRA configurations and backbones; LifeFuse-Mem adds a four-plastic/four-stable partition and token-level permanent routing, with protected-state fusion only in the controlled benchmark.Public benchmarks use standard online-memory mode without Phase-A checkpointing, stable-row restoration, or query-specific fusion; lifecycle metadata supervises routing during training.Source: §5.1 · p. 8

5.2. Hard Attribution Anti-Overwrite

• Ret. rises from 57.72% to 63.71% on Qwen3-4B and from 60.27% to 69.39% on SmolLM3-3B versus δ-Mem, while Ovr. falls in parallel.These acquisition-controlled metrics cover facts answered correctly after Phase A; Table 1 reports the benchmark results.Source: Table 1 · p. 9, §5.2 · pp. 9–10

5.3. Long-Memory Benchmarks

• Against δ-Mem, LifeFuse-Mem improves LoCoMo from 41.24% to 42.41% and MemoryAgentBench from 31.83% to 32.24% on Qwen3-4B, but category results vary.On SmolLM3-3B, LoCoMo rises from 30.65% to 38.30%, while MemoryAgentBench changes from 20.13% to 20.21%; neither public benchmark tests phase-aware anti-overwrite directly.Source: Table 2 · p. 9, §5.3 · p. 11

5.2. Hard Attribution Anti-Overwrite

• The controlled benchmark tests recovery of Phase-A permanent facts after conflicting Phase-B writes, with Ret. and Ovr. evaluated only on facts acquired after Phase A.Temp. is an auxiliary check of Phase-B behavior, and Ret. should be read alongside end-to-end Perm.Source: §5.2 · pp. 9–10

5.4. Ablation Studies

• Removing route supervision lowers Ret. from 63.71% to 58.68% on Qwen3-4B and from 69.39% to 59.45% on SmolLM3-3B, identifying it as the most reliable retention contributor.The ablation indicates rank partitioning alone is insufficient to make the subspaces lifecycle-aware.Source: Table 3 · p. 10, §5.4 · p. 11
• Removing stable rows or permanent retention loss weakens acquisition-controlled retention on both backbones, while hard-candidate loss has model-dependent effects and does not uniformly improve Ret. or Ovr.On Qwen3-4B, removing hard-candidate loss slightly improves Ret. and Ovr. but reduces Perm. from 20.12% to 19.28%.Source: §5.4 · pp. 11–12

• LifeFuse-Mem reduces acquisition-conditioned overwrite on the controlled Hard Attribution Anti-Overwrite benchmark through lifecycle-aware routing, subspace isolation, and protected readout.Source: §6 · p. 12
• Public-benchmark compatibility persists when protected-state fusion is unavailable, but gains vary across backbones and task categories.Source: §6 · p. 12
• Ablations identify route supervision as the most reliable source of retention gains, while stable rows, protected fusion, and auxiliary losses contribute more selectively.Source: §6 · p. 12
• Results support explicit lifecycle modeling as a diagnostic and architectural principle, while lifecycle inference from unlabeled histories remains future work.Source: §6 · p. 12