Different Agents, Shared Memory: Building Persistent Context with Neo4j

AI agents can remember conversations and tasks, but their context is often tied to a particular agent, model, or session. When we switch to another AI tool, we may need to explain the same things again. Decisions, supporting evidence, and the reasons behind them can also be lost.

This talk explores a broader question: What knowledge should we preserve, and how can we share it across different AI agents?

Rather than storing entire conversation histories, the focus is on preserving useful knowledge: facts, relationships, decisions, evidence, and relevant changes over time. Knowledge graphs provide a way to connect this information while keeping its context and sources visible.

The presentation introduces Neo4j as a shared knowledge store and shows how ChatGPT can access a Neo4j Aura database through MCP (Model Context Protocol). It also explores how multiple agents can work with the same persistent knowledge while maintaining independent conversations and reasoning.

A case study of Yabitsu Port, an underwater cultural heritage site on Tsushima Island, Japan, illustrates another important challenge: distinguishing verified facts from unverified claims and preserving the sources and uncertainty behind historical information.

Finally, the talk introduces SEPT, a proposed set of principles for shared knowledge:

Shared: Accessible across agents

Evolving: Knowledge changes over time

Plural: Multiple perspectives coexist

Traceable: Sources and reasons remain visible

The central message is simple: Knowledge should outlive the agent.

Note: The slides have been slightly revised after the presentation to clarify a few points and avoid potential misunderstandings.

Transcript

1. Tokyo AI (TAI) — Wednesday, 14 October 2026 — 19:30–20:00 (UTC+9) Different Agents, Shared Memory Building Persistent Context with Neo4j Koji Annoura Neo4j Ninja @kojiannoura koji@annoura.com Source: Conceptual illustration created by Koji Annoura with OpenAI ChatGPT, 2026.
2. Who Am I? ❖ Koji Annoura ❖ Fukuoka, Japan ❖ Neo4j Ninja ❖ Co-founder of Neo4j Users Group Tokyo ❖ Graph Community MVP Asia 2020 ❖ Building Practical Knowledge Graphs ❖ Cypher, GQL, SQL/PGQ ❖ Diver
3. Today’s Journey ❖ 1. Questions & Challenges ❖ 2. What Knowledge Should We Keep? ❖ 3. Sharing Knowledge with Neo4j ❖ 4. Case Study: Yabitsu Port ❖ 5. SEPT & Takeaways
4. 1. Questions & Challenges
5. How Many AI Tools Do You Use? ❖ AI Tools ❖ ChatGPT ❖ Claude ❖ Gemini ❖ Others ❖ Your Experience ❖ One AI tool? ❖ Two or more? ❖ Different tools for different tasks? Source: Conceptual illustration created by Koji Annoura with OpenAI ChatGPT, 2026.
6. Where Do You Keep Your Knowledge? ❖ Common Approaches ❖ AI conversation history ❖ Documents and notes ❖ Local les ❖ External databases ❖ A Question ❖ Can another AI reuse that knowledge? fi Source: Conceptual illustration created by Koji Annoura with OpenAI ChatGPT, 2026.
7. Explaining the Same Thing Again ❖ With Agent A ❖ Long discussion ❖ Requirements clari ed ❖ Decisions made ❖ Switching to Agent B ❖ No shared context ❖ Repeating explanations ❖ Starting over fi Source: Conceptual illustration created by Koji Annoura with OpenAI ChatGPT, 2026.
8. Why Did We Make That Decision? ❖ What Remains ❖ Final decision ❖ Selected solution ❖ What Gets Lost ❖ Alternatives considered ❖ Reasons for the choice ❖ Supporting evidence Source: Conceptual illustration created by Koji Annoura with OpenAI ChatGPT, 2026.
9. Different AI, Different Answers ❖ Agent A ❖ Recommends Solution A ❖ Based on one set of evidence ❖ Agent B ❖ Recommends Solution B ❖ Uses different assumptions ❖ The Challenge ❖ Compare evidence and reasoning Source: Conceptual illustration created by Koji Annoura with OpenAI ChatGPT, 2026.
10. Is This Knowledge Still Valid? ❖ Yesterday ❖ A fact was con rmed ❖ A decision was reasonable ❖ Today ❖ Requirements have changed ❖ New evidence is available ❖ Missing Information ❖ When and why it changed fi Source: Conceptual illustration created by Koji Annoura with OpenAI ChatGPT, 2026.
11. The Problem Is Bigger Than Memory ❖ Agent Memory ❖ Session-dependent ❖ Agent-speci ❖ Dif c cult to transfer ❖ Knowledge Management ❖ Shared context ❖ Persistent relationships ❖ Evidence and history fi fi Source: Conceptual illustration created by Koji Annoura with OpenAI ChatGPT, 2026.
12. 2. What Knowledge Should We Keep?
13. Do We Need to Remember Everything? ❖ Conversation History ❖ Questions and answers ❖ Trial and error ❖ Repeated explanations ❖ Knowledge Worth Keeping ❖ Con rmed facts ❖ Important relationships ❖ Decisions and reasons ❖ Sources and changes fi Source: Conceptual illustration created by Koji Annoura with OpenAI ChatGPT, 2026.
14. Small Knowledge, Big Value ❖ Simple Knowledge ❖ Entities ❖ Relationships ❖ Properties ❖ Reusable Context ❖ What was decided? ❖ Why was it decided? ❖ Where did the evidence come from? ❖ Key Point ❖ Small, explicit, reusable Source: Conceptual illustration created by Koji Annoura with OpenAI ChatGPT, 2026.
15. 3. Sharing Knowledge with Neo4j
16. Why a Graph? ❖ Facts Alone ❖ Project requirements ❖ Technology choices ❖ Supporting documents ❖ Connected Knowledge ❖ Decisions depend on requirements ❖ Evidence supports decisions ❖ Changes affect relationships Source: Conceptual illustration created by Koji Annoura with OpenAI ChatGPT, 2026.
17. What Is Neo4j? ❖ Graph Database ❖ Nodes ❖ Relationships ❖ Properties ❖ Working with Graphs ❖ Cypher query language ❖ Graph visualization ❖ Connected data exploration ❖ Our Use Case ❖ Persistent shared knowledge Source: Conceptual illustration created by Koji Annoura with OpenAI ChatGPT, 2026.
18. Three Ways to Get Started ❖ Try Neo4j ❖ AuraDB Free ❖ fully managed cloud database ❖ Neo4j for Desktop ❖ local development ❖ Neo4j Community Edition ❖ open source, self-managed ❖ Easy to start ❖ neo4j.com/download/ Source: Neo4j, neo4j.com/download/, accessed 10 Aug 2026.
19. Start with Neo4j Aura ❖ No credit card required ❖ Start free Source: Neo4j Aura Console, © Neo4j, Inc.
20. Create Your Account ❖ Create an account ❖ Enter basic details Source: Neo4j Aura Console, © Neo4j, Inc.
21. Connections Appear Immediately ❖ Data connects Source: Neo4j Aura Console, © Neo4j, Inc.
22. Choose Where Your Graph Lives ❖ Choose a region ❖ Start your database Source: Neo4j Aura Console, © Neo4j, Inc.
23. You Can Build a Graph with AI ❖ Describe what you need ❖ Build it step by step Source: Neo4j Aura Console, © Neo4j, Inc.
24. Explore the Graph Visually Source: Neo4j Aura Console, © Neo4j, Inc.
25. MCP for Aura ❖ Connect AI to your graph ❖ No local MCP server Screenshot: Neo4j Aura Console, © Neo4j, Inc., captured and redacted by Koji Annoura.
26. Connect ChatGPT to MCP for Aura ❖ ChatGPT ❖ Settings ❖ Plugins ❖ MCPs Screenshot: ChatGPT Work, © OpenAI, captured and redacted by Koji Annoura.
27. ChatGPT Queries Aura through MCP ❖ Connected to the Aura graph ❖ Who is in The Matrix? Screenshot: ChatGPT Work, © OpenAI, captured by Koji Annoura.
28. One Graph, Multiple Agents ❖ Different Agents ❖ Independent conversations ❖ Different models ❖ Different reasoning ❖ Shared Knowledge ❖ Common entities ❖ Persistent relationships ❖ Reusable context Source: Conceptual illustration created by Koji Annoura with OpenAI ChatGPT, 2026.
29. Connecting Agents with MCP ❖ Model Context Protocol ❖ Standardized tool access ❖ External data and services ❖ Agent-independent interface ❖ Neo4j Integration ❖ Query existing knowledge ❖ Add new knowledge ❖ Update relationships Source: Conceptual illustration created by Koji Annoura with OpenAI ChatGPT, 2026.
30. Example Workflow: Sharing Knowledge ❖ Agent A ❖ Discuss a technology decision ❖ Save the decision and its reasons ❖ Agent B ❖ Retrieve the same knowledge ❖ Add new evidence ❖ Back to Agent A ❖ Retrieve the updated knowledge Source: Conceptual illustration created by Koji Annoura with OpenAI ChatGPT, 2026.
31. 4. Case Study: Yabitsu Port
32. A Small Example: Yabitsu Port ❖ Yabitsu Port, Tsushima Island, Japan ❖ A small port in a narrow coastal inlet ❖ Stone structures on land and underwater ❖ Historical evidence remains at the site ❖ Many questions are still unanswered Map data © OpenStreetMap contributors
33. Hidden Underwater Heritage ❖ What Can We Learn from Underwater Remains? ❖ Stone structures beneath the water ❖ Dif cult to access and observe ❖ Evidence recorded through photographs and videos ❖ Historical context is often incomplete fi fi Photo © Koji Annoura, Yabitsu eld survey, September 2023
34. Knowledge Is Hidden Too ❖ Evidence Is Scattered Across Different Sources ❖ Photographs and survey videos ❖ Historical documents and reports ❖ Maps and geographic data ❖ Observations and interpretations ❖ Relationships are often missing fi Photo © Koji Annoura, Yabitsu eld survey, September 2023
35. When Was Yabitsu Port Built? ❖ Different Structures, Different Histories? ❖ Construction dates ❖ When were the stone walls built? ❖ Regional comparisons ❖ How do construction techniques compare with other regions? ❖ Differences within the site ❖ Stone construction varies by location. ❖ Multiple periods ❖ Was the port built, modi ed, or reused over time? fi fi Photo © Koji Annoura, Yabitsu eld survey, September 2023
36. ChatGPT's Response ❖ Japanese ❖ 室町時代（1336〜1573年）：朝鮮側の記録に、通信使の船が矢櫃に停泊したことが記され ています。 ❖ Translate ❖ Muromachi period (1336–1573): Korean historical records reportedly mention that Korean diplomatic envoys' ships stopped at Yabitsu Port.
37. Modeling Historical Evidence as a Knowledge Graph ❖ (:Event)-[:AT_PORT]->(:Port) ❖ (:Event)-[:DURING]->(:Period) ❖ (:HistoricalSource)-[:MENTIONS]->(:Event)
38. ChatGPT's Unverified Response ❖ Japanese ❖ 室町時代（1336〜1573年）：朝鮮側の記録に、通信使の船が矢櫃に停泊したことが記され ています。 ❖ English (Translate) ❖ Muromachi period (1336–1573): Korean historical records reportedly mention that Korean diplomatic envoys' ships stopped at Yabitsu Port.
39. Historical Evidence of Yabitsu Port ❖ Japanese ❖ 室町時代（1336 1573年）：中世朝鮮の記録には、通信使の船が矢櫃に停泊したとの記述 があるとされています。 ❖ English (Translation) ❖ Secondary sources suggest that medieval Korean records mention diplomatic envoys' fi ～ ships stopping at Yabitsu Port during the Muromachi period (1336–1573). The original records have not yet been veri ed.
40. fi Modeling Historical Evidence and Uncertainty ❖ The original historical source has not yet been veri ed. ❖ (:SecondarySource)-[:REPORTS]->(:Claim) ❖ (:Claim)-[:ABOUT]->(:Event) ❖ (:Event)-[:AT_PORT]->(:Port) ❖ (:Event)-[:DURING]->(:Period) ❖ (:Claim)-[:CITES]->(:PrimarySource)
41. Mapping Yabitsu Port ❖ Understanding the Physical Environment ❖ Water depth and seabed topography ❖ Shape of the coastal inlet ❖ Location of underwater structures ❖ Spatial relationships between features Bathymetric contour map created by Windy Network. Source: ARIUA (Asian Research Institute of Underwater Archaeology)
42. Surveying and Recording the Site ❖ Field Survey — September 16–17, 2023 ❖ Surveying structures on land and underwater ❖ Removing debris to improve visibility ❖ Recording damage and structural conditions ❖ Documenting observations from different locations fi Photo © Koji Annoura, Yabitsu eld survey, September 2023
43. Garbage Collection on the Shore ❖ We cleaned drifted waste from the shore. ❖ Plastic ❖ Fishing gear ❖ Other debris ❖ Cleaning also made the site easier to observe. fi Photo © Koji Annoura, Yabitsu eld survey, September 2023
44. Recording the Damage ❖ Some parts of the remains were damaged. ❖ We wanted to understand why. ❖ We took photos on land ❖ We took photos underwater ❖ We recorded the condition from different angles fi Photo © Koji Annoura, Yabitsu eld survey, September 2023
45. Who Built This Underwater Quay? ❖ 2018-08-05 ❖ Yabitsu No.A ❖ ARIUA Video: ARIUA (Association for Research Institute of Underwater Archaeology), YouTube
46. Who Built This Underwater Quay? ❖ 2018-08-05 ❖ Yabitsu No.B ❖ ARIUA Video: ARIUA (Association for Research Institute of Underwater Archaeology), YouTube
47. 5. SEPT & Takeaways
48. Knowledge Is More Than Facts ❖ Facts ❖ What is known? ❖ What was decided? ❖ Context ❖ Why was it decided? ❖ Who provided the evidence? ❖ History ❖ What changed? ❖ When did it change? Source: Conceptual illustration created by Koji Annoura with OpenAI ChatGPT, 2026.
49. SEPT: Principles for Shared Knowledge ❖ Shared ❖ Accessible across agents ❖ Evolving ❖ Knowledge changes over time ❖ Plural ❖ Multiple perspectives coexist ❖ Traceable ❖ Sources and reasons remain visible SEPT: A proposed framework for persistent, shared knowledge. Source: Conceptual illustration created by Koji Annoura with OpenAI ChatGPT, 2026.
50. What Should We Preserve? ❖ Keep ❖ Reusable facts and relationships ❖ Decisions and reasons ❖ Evidence and provenance ❖ Relevant history ❖ Not Everything ❖ Temporary questions ❖ Repeated explanations ❖ Unnecessary conversation details ❖ Principle ❖ Preserve meaning, not volume Source: Conceptual illustration created by Koji Annoura with OpenAI ChatGPT, 2026.
51. Knowledge Should Outlive the Agent ❖ Models Will Change ❖ New models ❖ New agents ❖ New tools ❖ Knowledge Should Remain ❖ Shared ❖ Persistent ❖ Reusable Source: Conceptual illustration created by Koji Annoura with OpenAI ChatGPT, 2026.
52. Key Takeaway ❖ Knowledge Should Outlive the Agent. ❖ Keep knowledge outside individual AI agents. ❖ Preserve essential facts, relationships, decisions, and evidence. ❖ Share and reuse knowledge across agents through Neo4j and MCP. ❖ Keep knowledge small, evolving, and traceable.
53. Thank You ❖ Thanks for listening! ❖ Koji Annoura ❖ https://annoura.com ❖ koji@annoura.com ❖ @kojiannoura