Skip to content

A-TMA: Decoupling State-Aware Memory Failures in Long-Term Agent Memory

🕒 Published (v1): 2026-07-02 09:28 UTC · Source: Arxiv · link

Why this paper was selected

Ghost memory state-coordination failures in long-term agent memory — concrete buildable fix

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

Long-term LLM agent memory fails when outdated, current, and transitional facts coexist without explicit state roles — a failure the authors term ghost memory. A-TMA is a state-aware overlay that wraps existing memory systems with bank-level supersession tracking, state-aligned retrieval, and labeled evidence presentation to QA. On the new LTP benchmark, pairing A-TMA with Graphiti/Zep raises conflict accuracy by 0.240 absolute; on LoCoMo, temporal F1 rises from 0.0295 to 0.1705.

Problem

Existing long-term memory systems (profile-based, temporal, graph-based) treat facts as timeless or rely on recency/timestamps alone. When user facts change (address, job, plan), old and new facts accumulate in the bank without explicit state roles. During retrieval, both states surface as equally "relevant," and the answer model cannot distinguish which state the query requests. A single final-QA accuracy score obscures where the failure occurs (bank, retrieval, or generation). No prior benchmark was specifically designed to stress this failure mode.

Method

A-TMA is a host-agnostic overlay that adds state semantics to three pipeline levels without replacing the underlying memory substrate.

Bank module — state maintenance via Sentry + Judge: Each record \(r_i = (x_i, t_i, s_i, L_i, z_i)\) carries status \(s_i \in \{\text{active, superseded, transition, unknown}\}\) and typed links (\(\text{supersedes}\), \(\text{superseded\_by}\), \(\text{evolves\_from}\), etc.). On write, a lightweight Sentry model (nomic-embed-text-v1.5 encoder with two normalized projection heads) scores candidate pairs on topic agreement \(a_{ij}\) and logical compatibility \(b_{ij}\): $\(A_j = \{i \in P_j : a_{ij} > \tau_\text{top} \;\wedge\; (b_{ij} < \tau_\text{log} \;\vee\; d_{ij} > \tau_\text{gap})\}\)$ Sentry is trained with a supervised dual-head objective on same-slot conflict/consistent/negative pairs. Flagged pairs are passed to a Judge (Qwen2.5-3B-Instruct, QLoRA SFT + GRPO) that returns a structured verdict and commits a supersession or transition link, marking the old record superseded while preserving it for historical queries.

Retrieval module — state-aligned evidence packet: A rule-based query profiler assigns \(v_q \in \{\text{current, historical, transition, neutral}\}\). The evidence pool \(C_q\) is formed by concatenating host seed results \(A_q\), relation-hop expansions \(R_q\) (following supersession/transition links view-conditionally), and semantic neighbor hops \(N_q\), then deduplicating. A stable pre-rank preserves host semantic ordering; an optional bounded controller re-ranks by state-match signal. The top-\(k\) packet \(E_q\) retains per-record source bucket, status, and relation type as trace metadata.

QA module — explicit state-role conditioning: Evidence is serialized with explicit labels (current memory, historical memory, transition memory, transition-linked memory) before the answer model, conditioning it to resolve according to \(v_q\) rather than collapsing all states into one response.

Benchmark — LTP (LoCoMo Temporal Plus): 10 profiles, 800 probes with paired old/current states and transition records; structurally validated and manually verified. Evaluated alongside LoCoMo (10 samples, 1,986 QA pairs) for long-conversation generalization.

Key Contributions

  • Formalizes ghost memory as a three-level (bank / retrieval / QA) state coordination failure, distinct from simple staleness or deletion problems.
  • Introduces A-TMA, a host-agnostic overlay with Sentry (lightweight dual-head encoder) and Judge (fine-tuned 3B LM with GRPO) for principled supersession commits.
  • Constructs LTP, a conflict-heavy benchmark targeting old/current state discrimination with decoupled bank, retrieval, and QA metrics.
  • Proposes decoupled evaluation protocol so that bank preservation, retrieval state-fit, and QA resolution are measured independently rather than collapsed into a single accuracy number.

Results

  • LTP conflict accuracy: Graphiti/Zep + A-TMA: 0.720 vs. Graphiti/Zep baseline 0.480 (+0.240 absolute).
  • LoCoMo temporal F1: Graphiti/Zep + A-TMA: 0.1705 vs. baseline 0.0295 (+0.1410).
  • LoCoMo average F1: Graphiti/Zep + A-TMA: 0.1556 vs. baseline 0.0809.
  • LTP subset (profile000): A-Mem + A-TMA reaches 0.775 conflict accuracy vs. AriadneMem (timestamped graph) at 0.438.
  • Retrieval evidence support vs. answer accuracy gap: Graphiti/Zep achieves 0.812 evidence support but only 0.425 conflict accuracy without A-TMA; adding A-TMA raises conflict accuracy to 0.725, illustrating that evidence presence alone is insufficient.
  • Gains are host-dependent; on LoCoMo sample0, AriadneMem remains stronger on F1 while A-TMA improves the Graphiti pair across F1, QA accuracy, evidence hit rate, and temporal F1.

Limitations

  • Gains are host-dependent; hosts that do not expose sufficient evidence structure show limited benefit.
  • LTP currently targets only current/historical query views; transition queries are used as evidence labels, not direct QA targets.
  • LTP has 10 profiles and 800 probes — relatively small scale; generalization to diverse real-world fact-change distributions is unverified.
  • The Sentry model requires labeled pair data (conflict/consistent/negative) that may be expensive to obtain for new domains.
  • The Judge (Qwen2.5-3B) introduces non-trivial inference cost on every write path; latency and throughput at scale are not analyzed.
  • The paper notes the "active tier training script" distinction without fully ablating the contribution of GRPO vs. SFT alone for the Judge.

Relevance to Agentic AI / LLM Agents

Long-term memory is increasingly core infrastructure for persistent LLM agents (task memory, user preference tracking, multi-session assistants), and ghost memory is a concrete failure mode that becomes more acute as agents are deployed over long horizons where user facts genuinely change. A-TMA's overlay design is practically significant because it is host-agnostic — it can be applied on top of Graphiti, Mem0, A-Mem, or other substrates without retraining the memory backbone, lowering adoption friction. The three-level diagnostic framework (bank / retrieval / QA) provides a principled vocabulary for agent memory evaluation that current single-score benchmarks lack, which is directly useful for teams building or auditing production memory pipelines. The LTP benchmark fills a gap between generic long-context QA (LoCoMo) and stress-testing state-change coordination, giving the community a concrete target for measuring memory reliability in agentic settings.