h t t p s : / / a r x i v . o r g / a b s / 2 6 0 9 . 1 2 4 3 6

Ingest: LifeFuse-Mem: Lifecycle-Aware State Fusion Against Temporary Overwriting

Type: kb/sources/types/ingest-report.md

Classification

Scientific paper: an arXiv preprint by Hanyu Zhao, Yuqian Feng, Zhenyu Song, Yuanchao Cheng, Yance Jiao, Tengfei Pan, and Li Du, affiliated with Chinese research institutions including the Beijing Academy of Artificial Intelligence. The authors propose a neural memory adapter and report a controlled benchmark, public-benchmark comparisons, and component ablations. This is author-reported experimental evidence; the captured representation is the main paper, without its separately referenced Appendix file.

Summary

LifeFuse-Mem studies temporary overwrite: a locally valid exception displaces a durable fact in compact neural memory. It trains a router using permanent/temporary episode labels and learns a coordinate basis with four plastic and four stable dimensions. Its controlled evaluation additionally supplies write-phase boundaries and query lifecycles, preserves the state before conflicting writes, and fuses scores from that checkpoint and a state with restored stable rows. Under this combined training and readout arrangement, retention among initially acquired conflicting facts rises over δ-Mem from 57.72% to 63.71% on Qwen3-4B and from 60.27% to 69.39% on SmolLM3-3B. Final accuracy across all permanent queries remains near 20%. Public LoCoMo and MemoryAgentBench runs omit checkpoint restoration and lifecycle-specific fusion; they show compatibility with ordinary memory use, with LoCoMo gains and mixed MemoryAgentBench category results. The most reusable contribution is the separation of acquisition failure from subsequent overwrite, rather than evidence for autonomous lifecycle discovery or a generally superior memory architecture.

Quotes

No source quotes have been retained yet.

Connections Found

The source supplies a neural-memory comparison for Retire, Redact, Supersede, And Relax Memory: lifecycle distinctions affect behavior only when writes or reads consume them. Here a temporary exception remains locally usable while earlier evidence is protected. That differs from permanent supersession and does not establish redaction, retirement, or revision operations. The retention gains apply to supervised routing plus phase-aware checkpoint readout, so they cannot directly validate lifecycle handling in a Markdown KB.

It also provides a bounded example for Learning inside a fixed decomposition inherits its mistakes. The coordinate basis and router are learned, while the two lifecycle classes, rank allocation, and controlled readout interface remain supplied choices. The experiment establishes improvement within this arrangement without comparing it against alternative memory representations or lifecycle ontologies.

Learning Claims (our opinion)

Training combines answer prediction after memory writes, lifecycle route supervision, candidate-ranking and retention losses, and regularization. Online adaptation then updates a compact associative state: plastic rows remain writable, while the predicted permanent-route weight controls stable-row erase and write strength. Route labels originate in episode metadata and are applied to valid write tokens. Removing route supervision produces the largest reported retention loss among the four component ablations: 5.03 percentage points on Qwen3-4B and 9.94 on SmolLM3-3B, with the remaining setup held fixed. This supports the value of that supplied training signal in this configuration; it does not show that all useful lifecycle information must be supplied explicitly.

In Commonplace terms, this is adaptation of distributed neural state under a designed update rule, not demonstrated theory refinement. Stable coordinates do not expose fact-level premises for diagnosis and selective repair. Learning can change the router and coordinate basis, but does not revise the permanent/temporary distinction, the four/four allocation, or the protocol that chooses which checkpoint to consult. The source therefore adds a concrete example of learning within a supplied lifecycle boundary. It leaves open whether a learner could infer and revise that boundary from ordinary histories; the controlled result does not answer that question.

Extractable Value

  1. Separate acquisition from retention in memory tests. Test the durable fact before introducing a conflicting temporary value, then report both retention on the acquired subset and final accuracy over all permanent queries. This diagnostic distinction can transfer across memory substrates, although its effectiveness for Commonplace remains untested. Retention and overwrite are complements here only because the candidate set contains the permanent answer and temporary alternatives. [quick-win]
  2. Test temporary exceptions separately from lasting replacements. The lifecycle note already requires operational lifecycle handling; this source makes the distinction between local validity and future authority concrete. A KB evaluation could test use of an exception during its scope and recovery of the durable rule afterward, without assuming neural subspace isolation is the relevant implementation. [experiment]
  3. Retain a qualified neural-memory design reference. Learned routing, stable-row protection, and checkpoint score fusion offer one implementation under explicit lifecycle supervision and known evaluation phases. The ablations favor route supervision most strongly; they do not establish that the rank partition or every auxiliary loss is necessary. [just-a-reference]

Limitations (our opinion)

The direct anti-overwrite evidence comes from a constructed 1,000-episode benchmark on two small backbones, sharing a rank-8 memory configuration and a 3,243-sample training set. The full method receives additional lifecycle supervision and uses privileged phase and query information at evaluation. The comparison therefore measures a compound intervention, rather than isolation alone. The no-route-supervision ablation varies the training signal; it does not remove the evaluation's knowledge of phases or query lifecycles.

The controlled permanent-query score assigns weight 0.70 to the pre-conflict checkpoint. A simpler explanation for part of the gain is access to that preserved evidence. The main paper does not supply the referenced fusion-weight sensitivity results or a matched checkpoint-only comparison sufficient to isolate the value of protected-state fusion. The stable-row ablation changes retention by only 0.92 percentage points on each backbone, much less than removing route supervision. These results do not establish that increasing capacity, replaying prior state, or choosing another representation could not solve the same task.

Conditional retention uses each method's initially acquired subset, whose membership may differ across methods. It is a useful diagnostic, but a higher value can coexist with weaker overall accuracy: SmolLM3-3B permanent-query accuracy falls from 19.95% to 19.65% relative to δ-Mem. Temporary-query accuracy also stays near 20%, limiting claims of robust local adaptation. The reported tables lack uncertainty estimates or seed variation, and this ingest neither inspected implementation code nor reproduced results. Public benchmarks omit the protected readout and cannot fill the direct anti-overwrite evidence gap. Neither benchmark setting demonstrates durable policy enforcement in deployed agents or lifecycle handling in file-backed KBs.

Update the evaluation questions in Retire, Redact, Supersede, And Relax Memory with one acquisition-controlled temporary-exception test that checks local exception use, later durable-rule recovery, and overall accuracy separately.