AMemOps: Evaluating Memory Updates in Long Audio Streams

1. INTRODUCTION

• A returning source may change pitch and level while its spatial position remains unchanged, so retrieval must update one attribute and preserve the other.The introduction identifies source recognition, history access and attribute updating as possible sources of retrieval errors.Source: §1 · p. 1
• Attention retrieves from stored observations, whereas delta-rule memories update key–value associations in fixed-size state; both depend on reliable source representations.Attention must select the latest matching occurrence, while recurrent updates require consistent source keys.Source: §1 · p. 1
• Existing benchmarks leave open how recording variation and obsolete acoustic observations jointly affect retrieval of a source’s latest attributes.Prior tasks cover explicit-key replacement, conversational updates and audio retrieval, but do not require recognizing a source across recordings.Source: §1 · p. 1
• AMemOps tests latest-attribute retrieval, unchanged-attribute retention and absent-source rejection without source labels.Its paired controls vary recordings, training identities and earlier attributes to examine their effects on updating.Source: §1 · p. 1

2. METHODS

• Streams combine nonoverlapping excerpts, distractors and queries with independently controlled gain and stereo pan for each sound class or instrument.Excerpts receive sampled gain and pan, then independent channel noise; panning preserves total event energy.Source: §2 · p. 1
• Queries request a source’s latest gain or pan, or “none” for an absent source, using answer-excluded query attributes to avoid revealing the answer.Separate gain and pan classifiers predict four values plus none; query frames update model memory but not ground-truth attributes.Source: §2 · p. 1
• GDN2 uses separate decay, erase and write gates to update key–value state, while matched attention uses causal softmax with recurrent models’ head dimensions and processing.The memory decays old contents, erases along the current key, writes a new association and reads from the updated state; KDA, GDN, no-erase and Mamba2-style variants provide controlled comparisons.Source: §2 · pp. 1–2
• Recording controls range from replaying one recording per source to reusing two or four recordings or using all recordings, while preserving timing and attributes.The NSynth familiarity setup trains on one instrument set and tests on both; instruments are “seen” when training includes other recordings of them.Source: §2 · p. 2
• Matched histories vary whether an earlier target matches or differs from the latest event while holding the latest event, query and answer fixed.The study tests 20- and 120-second delays; distractor load varies while counts, identities and durations remain fixed, with controls omitting or replaying the earlier target.Source: §2 · pp. 2–3
• Source-key projections edit memory state at the target’s final occurrence, with random and overlapping-source edits as norm-matched controls.The intervention applies S ← S − PkS and evaluates queries about bindings established before the edit that were not refreshed afterward.Source: §2 · p. 3

3. EXPERIMENTS

• The experiments vary recordings, training identities, earlier target attributes, retrieval delay and distractor load in controlled comparisons.Models are evaluated with paired stream and seed analyses; the identity controls keep each set’s test streams fixed across training conditions.Source: §3 · p. 3

3.1. Acoustic processing and recording variation

• 21.4 pp: matching recurrent heads and processing improves NSynth attention accuracy, while convolution helps retrieval across large pitch differences.On original recordings, GDN2 leads matched attention on updates by 12.1 pp with RoPE and 17.0 pp without; fixed recordings improve matched RoPE attention’s gain and pan update accuracy by 20.6 and 20.3 pp.Source: §3.1 · pp. 3–4, Figure 2 · p. 3

3.2. Seen and unseen instruments

• Training on other recordings of test instruments increases GDN2’s update advantage over matched RoPE attention by 13.4 pp.The increase is 9.6 pp for gain and 16.8 pp for pan; fixed recordings reduce the overall increase to 6.2 pp.Source: Figure 2 · p. 3, §3.2 · p. 4

3.3. Conflicting history at long retrieval delays

• Against these baselines, conflicting history reduces accuracy by 20.7 pp [12.0, 29.0] and 21.2 pp [14.1, 28.5], while consistent history adds 14.8 and 11.6 pp.At 120 s with four distractors, previous values account for 78% of no-RoPE attention’s update errors and 23% of GDN2’s; the four pre-specified interactions show no clear increase in GDN2’s update advantage with delay or load.Source: §3.3 · p. 4

3.4. Erasure, capacity and selective access

• With fixed recordings, GDN2 beats no-erase on updates by 11.3 pp on Set 2 and 9.5 pp on Set 3.The advantage is 5.5 pp on repeated-update ESC-50 streams; widening preserves the update advantage, and source-key edits outperform random edits on target loss and other-source errors.Source: §3.4 · p. 4, Figure 4 · p. 4

• AMemOps tests current-attribute retrieval and unchanged-attribute preservation, using paired controls to examine effects of acoustic variation, familiarity, conflicting observations and erasure.Source: §4 · p. 4