Source-linked AI summary
When Errors Become Memories: Causal Pathway Tracing in Multi-Turn Memory-Augmented LLMs
Shuyao Xiao, Shengling Wang, Xuan Chen, Ke Chao, Ming Cui, Feifei Qian, Fanlin Meng, Chaoyang Mei, Chaoyong Jiang, Qi Ouyang, Junxi Yi
TL;DR
Long-term memory can turn localized LLM errors into persistent cross-turn effects, while existing evaluations provide limited evidence about how those errors propagate through evolving questions, responses, and memory states. The paper models this process with a structural causal model and four counterfactual trajectories that intervene on memory updating and question feedback. Experiments find generally decaying influence, more persistent memory-pathway effects, and latent error traces beyond natural responses.
Problem
Existing research mainly evaluates whether memory systems retain and use information, leaving cross-turn error propagation across responses, memory states, and subsequent questions limitedly understood.
Method
The paper uses a dynamic SCM and four counterfactual trajectories to intervene on memory updating and question feedback and measure four downstream error outcomes.
Results
Error influence generally decays with interaction distance, while the memory-update pathway has more persistent effects than question feedback and latent traces can remain after natural responses recover.
Takeaways & Limitations
Pathway-guided restoration indicates that repairing memory is more effective than repairing question feedback alone, while jointly repairing both pathways nearly eliminates residual propagation.
Takeaways & Limitations
The causal model adopts a first-order Markov assumption in which variables depend only on the preceding round and current-round variables.
Abstract
from arXiv · showhide
Long-term memory enables large language models (LLMs) to preserve and reuse information across interactions, but it can also turn localized errors into persistent risks. Existing work mainly evaluates whether memory systems store and retrieve information correctly, leaving limited understanding of how errors propagate across responses, memory states, and future interactions. We propose a structural causal model (SCM)-based framework for cross-turn error propagation in memory-augmented LLMs. We model user questions, model responses, and memory states as a dynamic causal process, and identify two entry pathways: internal memory updating and external question feedback. By intervening on these pathways, we construct four counterfactual trajectories and quantify their downstream effects and interaction. Error influence is evaluated at four levels: memory retention, natural responses, targeted diagnostic probing, and probability-level error preference. Experiments show that error influence generally decays with interaction distance, while the memory-update pathway contributes more persistent effects than question feedback; latent errors may remain even after disappearing from natural responses. Propagation patterns also vary across memory categories and memory mechanisms. Pathway-guided restoration further validates this decomposition: Question Repair reduces residual error by 27.5%, Memory Repair by 70.2%, and Joint Repair by 98.3%, nearly eliminating residual propagation.
1 Introduction
Long-term memory can preserve and reuse information across sessions, but it can also let response errors persist and influence later interactions. This paper addresses the limited understanding of these cross-turn pathways with causal interventions and evaluates error influence across multiple manifestations.
- Motivation: Persistent memory can incorporate an incorrect response into later memory states, allowing the error to be repeatedly retrieved and reused.Such enduring effects may matter in health management, financial planning, and task execution.
- Research gap: Existing memory research mainly evaluates retention, retrieval, updating, and downstream task performance rather than dynamic error propagation across turns.The gap concerns how responses, memory states, and subsequent questions evolve together.
- Approach: The framework models questions, responses, and memory states jointly in an SCM and intervenes on internal memory updating and external question feedback.It measures memory retention, natural-response occurrence, targeted re-elicitation, and generation-probability shifts.
- Findings: Error influence generally decreases with interaction distance, while the internal memory pathway produces more persistent downstream effects than question feedback.Errors can remain as latent memory traces or generation-preference shifts after disappearing from natural responses.
- Restoration: Restoring both the corrupted memory state and the error-affected question trajectory nearly eliminates residual error propagation.The comparison indicates that memory restoration is more effective than correcting the question trajectory alone.
2 Related Work
Prior work has improved long-term memory architectures and evaluated their ability to retain and use historical information. Cross-turn causal propagation of errors across questions, responses, and memory states remains comparatively underexplored.
- Long-Term Memory-Augmented LLMs: Long-term memory systems support multi-session dialogue, information reuse, and extended-context processing through storage, retrieval, and memory management.Examples include MemoryBank, MemGPT, and LongMem.
- Long-Term Memory-Augmented LLMs: Most prior memory research prioritizes more effective memory architectures or improved downstream performance over multi-round information propagation.This leaves the dynamics of error persistence and spread less examined.
- Memory Evaluation in LLM Systems: Recent benchmarks assess retrieval, cross-session reasoning, temporal reasoning, updating, abstention, and related memory capabilities in interactive settings.These benchmarks extend evaluation beyond static long-context question answering.
- Causal Analysis of LLM Behavior: Causal studies have localized internal computations within single generation processes using interventions such as causal tracing, ROME, and activation patching.Their typical focus is on layers, token positions, activations, or other internal variables.
- Causal Analysis of LLM Behavior: Cross-turn system dynamics remain poorly understood because questions, responses, and memory states evolve jointly across multiple error pathways.The contribution of these pathways to long-term error propagation is the unresolved issue.
3 Structural Causal Modeling
The paper represents multi-turn memory-agent interaction as a first-order Markov SCM over questions, responses, and memory states. It distinguishes internal memory updating from external question feedback as two routes by which an erroneous response affects later interactions.
- Dynamic SCM: The SCM indexes each interaction round with Qr, Ar, and Mr denoting the user question, generated response, and available memory state.The formulation abstracts common interaction and memory-update processes in long-term memory systems.
- Dynamic SCM: A first-order Markov assumption makes each round depend on the preceding round and current-round variables while preserving mediated influence from earlier interactions.This provides a tractable abstraction of question evolution, response generation, and memory updating.
- Causal relationships: Memory persistence and update edges model how stored information carries forward and how queries and responses enter the subsequent memory state.The relevant relationships include Mr−1 → Mr, Qr−1 → Mr, and Ar−1 → Mr.
- Causal relationships: Response generation is conditioned jointly on the current question and the memory state available at that round.The graph represents this with Qr → Ar and Mr → Ar.
- Error pathways: The memory-update pathway changes the internal state, whereas the question-feedback pathway changes the external interaction context.The pathways are Ar−1 → Mr → Ar and Ar−1 → Qr → Ar, respectively.
4 Causal Intervention and Pathway Effects
The framework separates error entry through memory updating and question feedback, then compares four counterfactual trajectories as they evolve naturally. It measures propagation across memory, responses, targeted probes, and generation preferences using pathway-specific causal contrasts.
- The two successor mechanisms—memory updating and question feedback—form a 2 × 2 intervention design with four counterfactual trajectories.Tcc is fully clean; Tec uses memory updating only, Tce uses question feedback only, and Tee uses both mechanisms.
- At the injection round, matched correct and erroneous responses are separately passed to the memory updater, producing clean and erroneous memory states.
- After intervention, trajectories evolve independently with their assigned memory states, dialogue histories, and previous responses, enabling downstream causal contrasts.
- 4.1 Error-Propagation Outcomes: Propagation is evaluated through memory retention, natural responses, targeted diagnostic probing, and probability-level error preference across interaction distance k.The first three outcomes detect retained, explicit, or directly elicited errors; probability preference captures latent shifts in generation preference.
- 4.2 Pathway Effects: The total, memory-update, question-feedback, and interaction effects quantify overall, pathway-specific, and non-additive downstream influence.The interaction contrast is E[Yee(k)] − E[Yec(k)] − E[Yce(k)] + E[Ycc(k)], with positive and negative values indicating super-additive and sub-additive effects.
5 Experimental Evaluation
Experiments trace injected errors across models, outcomes, memory categories, and mechanisms using counterfactual pathway interventions. Error effects generally decay with distance, but memory updating produces more persistent traces and pathway-guided repair is most effective when both memory and dialogue are restored.
- The study constructs 144 interactions across six memory categories, each with 15 turns, injecting one target error at turn 3 and tracing it over 12 subsequent turns.
- Experiments evaluate natural responses, memory states, controlled probes, and probability-level error preference across Qwen2.5-7B-Instruct and OLMo3-7B-Instruct.
- Both models show strong short-range propagation followed by gradual decay, while memory-state and controlled-probe outcomes preserve longer-lasting traces than natural responses.
- At k = 1, memory-state total effects reach 1.0 for both models, while probability-level effects are 2.726 for Qwen2.5-7B-Instruct and 1.646 for OLMo3-7B-Instruct.
- The memory-update pathway contributes more strongly to memory, controlled-probe, and probability-level outcomes; interactions are mostly negative or near zero.
- Propagation varies by memory category and mechanism: preference errors persist most for Qwen2.5-7B-Instruct, plan errors for OLMo3-7B-Instruct, and episodic retrieval retains high memory-state effects over long distances.
- Memory Repair reduces residual MRE by 70.2%, compared with 27.5% for Question Repair, while Joint Repair reduces residual propagation by 98.3%.
6 Conclusion
The paper presents a causal framework for tracing cross-turn error propagation in long-term memory-augmented LLMs. Experiments show pathway-specific persistence, latent residual errors, and stronger restoration from repairing memory states than question feedback alone.
- The framework decomposes total error propagation into memory-update, question-feedback, and interaction components.
- Experiments across models, memory categories, and mechanisms show that errors can persist across multiple turns, with memory updates typically contributing more strongly.
- Latent errors can remain after disappearing from observable responses, while repairing both pathways nearly eliminates residual propagation.