AMU:Admission and Memory Update for Personalized Conversations---Structured Memory with SLM Guided Control

• Long-horizon assistants need stable, evolving user understanding, while conversational inputs mix reusable facts with transient, repeated, outdated, or weakly relevant information.When such inputs enter memory without sufficient control, they can accumulate as noisy, redundant, or outdated entries.Source: §1 · pp. 1–2
• Retrieval filtering cannot remove low-value records already written, and consolidation may still operate over a noisy candidate pool.The paper identifies admission control and relation-aware updates as coupled writing-time capabilities for long-term personalization.Source: §1 · p. 2
• AMU treats memory construction as a controlled writing process rather than passive accumulation before downstream retrieval.It targets low-value or non-personal information entering long-term memory and repeated or evolving states creating redundant or outdated records.Source: §1 · p. 2
• Structured admission filters low-value or non-personal inputs, while an SLM classifies admitted records as separate, duplicate, or updates to fuse.This couples decisions about what enters memory with how an admitted record interacts with existing memories.Source: §1 · p. 2

Personalized Memory

• Personalization studies commonly condition generation on personas, profiles, or dialogue histories, while admission and updating before retrieval receive less explicit study.AMU focuses on writing-time control, including whether an admitted memory is duplicate, an update, or separate.Source: §2 · pp. 2–3
• Persistent-memory systems include explicit memory operations, adaptive admission and conflict-aware maintenance, and structured linking with dynamic evolution.The cited examples are Mem0, A-MAC, and A-MEM, respectively.Source: §2 · p. 3

Retrieval-Augmented Memory

• Retrieval-augmented methods retrieve and use stored information at inference time, but do not fully address keeping personalized memory stores clean before retrieval.AMU complements these methods with structured filtering during memory writing and retrieval.Source: §2 · p. 3

Memory Construction and Management

• Agent methods use reasoning traces, feedback, or memory-like states, whereas personalized memory management must decide whether user utterances are stored, discarded, or used to update existing memories.AMU makes these relation judgments explicit before writing records back to memory.Source: §2 · p. 3

3.1 AMU Framework

• AMU converts user inputs into typed records, using structured fields to guide memory processing, candidate selection, storage relations, and retrieval.It decouples memory management from the response LLM, allowing the maintained store to be reused across different LLM backends.Source: §3.1 · pp. 3–4, Figure 2 · p. 4

3.2 Structured Memory Filtering

• Records contain aspect, intent, reusable-personal status, entities, an affect-neutral summary, sentiment, and confidence.Aspect captures what a memory is about; intent captures how the user presents it. Admission applies a validity gate and confidence threshold, with confidence elicited through a rubric rather than treated as a calibrated probability.Source: §3.2 · pp. 4–5
• Structured labels narrow candidate memories, with compatible intents under the same aspect considered when no exact aspect-intent match exists.Embedding similarity then restricts SLM judgments to plausible duplicate or update candidates.Source: §3.2 · p. 5

3.3 SLM-Guided Storage Management

• The SLM classifies each candidate as duplicate, update, or separate, and deterministic actions discard duplicates, insert separate records, or fuse updates.Only update candidates enter fusion; separate candidates remain unchanged, duplicates are excluded, and old update memories are removed.Source: §3.3 · p. 5

3.4 Structured Storage and RAG Retrieval

• Structured fields filter retrieval candidates before vector search, and retrieved memories are injected as plain-text context for RAG response generation.A retrieval-time threshold suppresses weakly related memories; it is distinct from the writing-time threshold used for duplicate and update detection.Source: §3.4 · pp. 5–6

4.1 Dataset and Experiment Setup

• The benchmark contains 20 theme-based sessions, each with 50 memory-writing turns and 10 evaluation queries, for 1,000 turns and 200 queries overall.Sessions simulate one to two months of interactions and include stable preferences, evolving attitudes, repeated information, and preference or plan updates.Source: §4.1 · p. 6

4.2 Manual Validation of Silver Labels

• Independent annotations found mean validity rates of 89.3% for action labels and 94.1% for query-memory alignment.The review sampled 600 of 1,000 writing turns and 100 of 200 evaluation queries; these rates measure annotation validity, not inter-annotator agreement.Source: §4.2 · p. 6, Table 1 · p. 6

4.3 Experimental Results

• In the controlled setting, AMU achieves 37.33 P@3, 63.50 R@3, 44.47 F1@3, and 57.50 HIT@3, with the lowest Red.@3, 2.83.Against A-MEM, AMU raises R@3 from 56.90 to 63.50 and F1@3 from 40.36 to 44.47, while Red.@3 is 2.83 versus 2.88.Source: §4.3 · p. 7

4.4 Component Verification

• Removing either structured filtering or SLM storage management weakens AMU; without SLM management, F1@3 and Action Acc. drop more substantially.The no-SLM variant's lower Redundancy@3 coincides with poor retrieval coverage and unreliable writing decisions, supporting the complementary roles of filtering and relation judgment.Source: Table 3 · p. 8, §4.4 · p. 8

5 Analysis

• The analysis relates writing-action accuracy to retrieval-side behavior as controller size changes.Source: §5 · p. 8
• After 4B, Action Acc. gains become limited, while F1@3 and Red.@3 do not improve proportionally.Retrieval also depends on embedding similarity, thresholds, and top-k ranking, so stronger writing-action prediction does not necessarily improve retrieval-side performance.Source: §5 · p. 8
• Qwen3.5-2B is sufficient for AMU in this setting without relying on a large response LLM for memory control.Controllers larger than 0.8B improve Action Acc., especially relative to 0.8B models.Source: §5 · p. 8

6 Conclusion

• AMU combines structured memory filtering with SLM-guided storage management for personalized conversations.Source: §6 · p. 8
• The evaluation uses a controlled memory-writing and retrieval benchmark to measure retrieval of expected memory slots while avoiding redundant memories.Source: §6 · p. 8
• The reported results support combining the two components to maintain cleaner personalized memories before downstream retrieval and response generation.Source: §6 · p. 8

Limitations of the Work

• Running-time results come from a single RTX 4070 local evaluation setting, not a large-scale serving environment.Larger controllers may run faster on stronger accelerators or optimized inference infrastructure.Source: Limitations of the Work · p. 8
• AMU's effectiveness depends on the chosen SLM's ability to follow structured prompts, produce stable fields, and distinguish duplicate, update, and separate relations.Source: Limitations of the Work · p. 8
• Very small or poorly aligned SLMs may produce unstable structures or unreliable relation judgments, affecting memory writing and downstream retrieval.Source: Limitations of the Work · p. 8

Ethics Statement

• The benchmark adapts LoCoMo-style long-term conversations for controlled evaluation without relying on real user conversations or personally identifiable private data.AMU-specific writing-action labels, memory identifiers, update links, and retrieval targets are created for the evaluation.Source: Ethics Statement · pp. 8–9
• Users should be informed when memory is enabled and be able to inspect, edit, or delete stored memories.Structured records allow selected entries to be removed or updated upon requests for deletion or correction.Source: Ethics Statement · p. 9

A Additional Analysis

• The appendix provides additional analyses of AMU's behavior under different settings.These are supplementary analyses beyond the primary evaluation.Source: A Additional Analysis · p. 12
• The analyses examine downstream LLM use, writing- and retrieval-time threshold sensitivity, embedding models, and SLM backbones.Source: A Additional Analysis · p. 12
• The appendix also analyzes session mixing with memories from multiple sessions.Source: A Additional Analysis · p. 12

A.1 Downstream LLM Judge Evaluation

• AMU receives the highest overall score in a small-scale response-level evaluation, while No Memory lacks relevant user information and Full Compressed Memory yields only modest improvement.The analysis is a supplement, not a full end-to-end personalization benchmark.Source: Table 5 · p. 12, §A.1 · p. 12

A.2 Threshold Sensitivity

• With τr fixed at 0.55, τs = 0.60 provides the highest Action Acc. and a balanced trade-off; higher τs reduces Time/Turn but can increase redundancy.Smaller τs retains more candidates for SLM-guided management and increases average processing time per turn.Source: Table 6 · p. 12, §A.2 · p. 12
• With τs fixed at 0.60, increasing τr improves P@3 and Redundancy@3 but reduces R@3 and HIT@3, exposing a retrieval precision-recall trade-off.The default τr = 0.55 preserves higher recall and HIT@3 while maintaining low redundancy.Source: Table 7 · p. 12, §A.2 · p. 12

A.3 Embedding Model Comparison

• The three tested embedding models produce close retrieval and writing-related metrics, and changing encoders does not substantially alter AMU's behavior.The same embedding model is used for writing-time candidate filtering and retrieval in each setting.Source: §A.3 · pp. 12–13

A.4 SLM Choice for Storage Management

• Changing the SLM backbone causes moderate differences but does not substantially change AMU's behavior; larger models generally perform better, especially on writing-action accuracy and retrieval metrics.Differences across most settings remain within a limited range.Source: §A.4 · p. 13

A.5 Session Mixing Analysis

• Retrieval metrics stay close across single-session and mixed-session memory stores, with no consistent degradation as the number of sessions increases.The analysis averages five random draws for single-, two-, and three-session stores.Source: §A.5 · p. 13, Table 8 · p. 14