Introducing Redis Agent Memory Server

Transcript

1. ⓒ 2026 Redis Ltd. All rights reserved. 1 Introducing Redis Agent Memory Server Raphael De Lio and Samuel Agbede
2. ⓒ 2026 Redis Ltd. All rights reserved. 2 Agenda • Demo • Context • Recurring Memory Patterns • Introducing Agentic Memory Server • Next steps
3. ⓒ 2026 Redis Ltd. All rights reserved. 3 Demo time
4. ⓒ 2026 Redis Ltd. All rights reserved. 4 Memory is not just storage It’s learning over time
5. ⓒ 2026 Redis Ltd. All rights reserved. 5
6. ⓒ 2026 Redis Ltd. All rights reserved. 6
7. ⓒ 2026 Redis Ltd. All rights reserved. 7 Not just chat history • User Preferences • Goals • Past actions • Learned knowledge • What worked
8. State/ history RAG Memory Prompt engineering Structured outputs User preferences User input User chat history Unstructured data storage Agents need the right context. LLM Goals Learned Knowledge
9. ⓒ 2026 Redis Ltd. All rights reserved. 9 Memory management strategies Periodically condense old turns into brief summaries ✅Retains longer-range knowledge within context limits 👎Relies on summary quality – additional compute Sequential (naive) From simple to smart Keep only the last N messages; drop the rest ✅Constant cost & latency 👎Forgets earlier details Feed entire chat history to the model every turn ✅Zero engineering effort 👎Context explosions → slow/expensive Sliding Window Summarization Extract & index memories. Fetch relevant pieces when possible. ✅Scales to unlimited history; adaptive recall 👎Relies on extraction quality Retrieval-based
10. ⓒ 2026 Redis Ltd. All rights reserved. 10 Recurring Challenges Memory Acquisition Memory Management Memory Retrieval
11. 11 ⓒ 2026 Redis Ltd. All rights reserved. Redis Agent Memory Server RESTful API server for managing short term and long term memory automatically Short-term memory • Automatic summarization • Configurable window sizes for recent messages Long-term memory • Search for relevant memories • Extract topic & named entity recognition • Namespace support for proper isolation Agent Memory Server User input AI app Short-term memory management Long-term memory management LLMs https://github.com/redis-developer/agent-memory-server
12. ⓒ 2026 Redis Ltd. All rights reserved. 12 Who decides what memory to extract? Hint: there is no one-size-fits-all answer
13. ⓒ 2026 Redis Ltd. All rights reserved. 13 Agent Memory Server Memory Integration Patterns Pattern Control Best For 🤖 LLM-Driven LLM decides Conversational agents, chatbots 📝 Code-Driven Your code decides Applications, workflows 🔄 Background Automatic extraction Learning systems
14. ⓒ 2026 Redis Ltd. All rights reserved. 14 What memory type gets extracted? Hint: there is no one-size-fits-all answer
15. ⓒ 2026 Redis Ltd. All rights reserved. 15 Agent Memory Server What should we remember? • Facts (default) • Summaries • Preferences • Custom
16. ⓒ 2026 Redis Ltd. All rights reserved. 16 How do we handle stale memories? Hint: there is no one-size-fits-all answer
17. ⓒ 2026 Redis Ltd. All rights reserved. 17 Memory Forgetting & Decay Strategies • Recency scoring: [ranking] ◦ Soft decay: freshness & novelty • Forgetting: [hard delete] ◦ age based ◦ Inactivity ◦ Budget • Summarization [Condensation]
18. ⓒ 2026 Redis Ltd. All rights reserved. 18 RECAP
19. ⓒ 2026 Redis Ltd. All rights reserved. 19 Memory is not just storage It’s learning over time
20. ⓒ 2026 Redis Ltd. All rights reserved. 20 Agents don’t fail alone because they can’t think They fail when they can’t remember
21. ⓒ 2026 Redis Ltd. All rights reserved. 21 Thank you very much!
22. ⓒ 2026 Redis Ltd. All rights reserved. 22 Raphael De Lio Samuel Agbede Slides and Demo *
23. ⓒ 2024 Redis Ltd. All rights reserved. 23 RECAP: Agent Memory Server feature comparison Redis Agent Memory Server Google Vertex MemoryBank AWS AgentCore Memory Auto-generate memories Yes, with choice of LLM and support for scopes, topics Yes, uses Gemini Yes, uses Bedrock models- Claude, Llama etc. Advanced retrieval Yes, Combine similarity, with recency, freshness, metadata etc. Limited, Similarity Search only (reference) Limited, Similarity Search only (reference) Decay and Freshness Yes, Configuring “forgetting” rules based on age, recency, budget etc. at namespace/userid level. Limited, TTL only Limited, TTL only Working memory mgmt Yes, incl. automatic expiration, tiering to/from long-term memory, No. Requires separate “Session” store and triggering generation Yes, includes automatic short-term (session-scoped) and long-term memory Cloud choice Yes, Redis available on all 3 clouds and on-prem. Memory Server is Open source today (w/ service coming) No, GCP only No, AWS only