cs.AIMar 1, 2026

Semantic XPath: Structured Agentic Memory Access for Conversational AI

Authors: Yifan Simon Liu, Ruifan Wu, Liam Gallagher, Jiazhou Liang, Armin Toroghi, Scott Sanner

Organizations: University of Toronto, Canada · Vector Institute of Artificial Intelligence, Toronto, Canada

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Conversational AI (ConvAI) agents increasingly maintain structured memory to support long-term, task-oriented interactions. In-context memory approaches append the growing history to the model input, which scales poorly under context-window limits. RAG-based methods retrieve request-relevant information, but most assume flat memory collections and ignore structure. We propose Semantic XPath, a tree-structured memory module to access and update structured conversational memory. Semantic XPath improves performance over flat-RAG baselines by 176.7% while using only 9.1% of the tokens required by in-context memory. We also introduce SemanticXPath Chat, an end-to-end ConvAI demo system that visualizes the structured memory and query execution details. Overall, this paper demonstrates a candidate for the next generation of long-term, task-oriented ConvAI systems built on structured memory.

Figures & tables

Figure 1: A travel-planning ConvAI agent for a 3-day ACL conference trip in San Diego. Left: in-context memory appends the full conversation. Middle: flat RAG retrieves from a flattened collection of memory items. Right: Semantic XPath retrieves and updates the relevant substructure from structured memory.
Figure 2: End-to-end Semantic XPath pipeline on an ACL 2026 trip plan illustrated in both the conversation view (blue) and the memory view (grey). Step 1: The user requests to “add a coffee break on the day that is packed with conference sessions.” is translated into a Semantic XPath query based on the memory schema. Step 2: The query is executed with structural matching and semantic relevance scoring, selecting Day 2 . Step 3: The retrieved substructure is passed to downstream generation to insert a coffee-break POI and respond to the user.
Figure 3: SemanticXPath Chat demonstration 1 . Left: Conversation view, where the user asks to add a coffee break on the day packed with conference sessions . Middle: Memory view, highlighting the retrieved Day 2 and the newly inserted coffee break. Right: Execution view, showing step-by-step query execution and scoring details.
Figure 4: Example of structured memory for three ConvAI scenarios: Travel Itinerary, To Do List, and Meal Kit Recommendation, with representative user requests covering Memory Read, Memory Write, and Memory Delete.
Figure 5: Single-turn evaluation across three methods. We report average pass rate ( left ) and token usage ( right ) averaged over three domains.
Figure 6: Multi-turn evaluation across three methods. We report average pass rate ( left ) and token usage ( right ) averaged over three domains.
Figure 7: Token usage vs. number of turns in a multi-turn ConvAI setting. Our proposed Semantic XPath maintains stable token consumption across turns, whereas the in-context baseline exhibits steadily increasing token usage as the conversation history grows.
Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

ModelMethodScorerItineraryTo Do ListMeal KitAvg.
   PassTokensPassTokensPassTokensPassTokens
GPT-5 miniIn-context–55.012,13680.05,44475.061,06970.026,216
Flat RAG–25.03,13740.02,72225.03,89230.03,250
GPT-5 miniSemantic XPathSem75.05,54864.05,93760.07,01366.36,166
 Semantic XPathEntail85.05,16184.05,18480.06,72883.05,691
Gemini 3 FlashIn-context–95.012,84672.010,46570.052,84979.025,387
Appendix
Table 1: Single-turn evaluation across methods. Sem denotes the Semantic Similarity scorer and Entail denotes the Entailment scorer for Semantic XPath . We report average pass rate and token usage with GPT-5 mini and Gemini 3 Flash as backbones across three domains: Travel Itinerary, To Do List, and Meal Kit Recommendation.
ModelMethodScorerItineraryTo Do ListMeal KitAvg.
   PassTokensPassTokensPassTokensPassTokens
GPT-5 miniIn-context–90.028,34270.020,39880.0158,55280.069,097
Flat RAG–20.03,25050.03,04310.03,68626.73,327
GPT-5 miniSemantic XPathSem50.05,664100.06,13260.07,08770.06,294
 Semantic XPathEntail80.05,820100.05,64970.06,07983.35,850
Gemini 3 FlashIn-context–60.031,63970.023,60240.0174,56756.776,603
Appendix
Table 2: Multi-turn evaluation across methods. Sem denotes the Semantic Similarity scorer and Entail denotes the Entailment scorer for Semantic XPath . We report average pass rate and token usage with GPT-5 mini and Gemini 3 Flash as backbones across three domains: Travel Itinerary, To Do List, and Meal Kit Recommendation.
More in this topic

Conversational Memory

Conversational memory equips AI assistants to retain, update, and retrieve useful information from past interactions for coherent, personalized dialogue.

See all 71 papers

Explore similar work

CardsList

  1. APEX-MEM: Agentic Semi-Structured Memory with Temporal Reasoning for Long-Term Conversational AI

    Apr 15, 2026Pratyay Banerjee, Masud Moshtaghi, Shivashankar Subramanian +2Long-Term Conversational MemoryGraph-Based Agent Memory

    Amazon, AGI, Sunnyvale, USA

  2. GRAVITY: Architecture-Agnostic Structured Anchoring for Long-Horizon Conversational Memory

    May 3, 2026Yushi Sun, Bowen Cao, Dong Fang +2LLM Agent MemoryPersistent Memory for Language Models

    LIGHTSPEED, Shenzhen, China · The Chinese University of Hong Kong, Hong Kong, China

  3. A Graph-Native Bitemporal Memory Store for Conversational AI Agents

    Jul 29, 2026Alp Niksarli, Gopesh BahetiConversational MemoryLLM Agent Memory

    Davidson College