Uploaded March 2026 | Updated November 2026, 5 days ago
github.com/supermemoryai/openclaw-supermemory
https://x.com/DhravyaShah/status/2023630749065228364
Timestamps
0:00 - The Evolution of Supermemory: Dhravya explains how his project started as a consumer bookmarking/note-taking app and evolved into a core infrastructure layer for LLM memory, focusing on user profiles, temporal reasoning, and knowledge graphs that avoid the inefficiencies of traditional triplets.
8:26 - The Problem with Current Memory Approaches: Dhravya critiques how OpenClaw and similar tools use static Markdown files and agentic tools for retrieval, which often leads to missing context, lack of freshness, and inconsistent performance. He proposes a hooks-based approach that dynamically injects relevant context.
16:32 - Benchmarks and Evaluation: The discussion covers the state of memory benchmarks like LongMemEval and LoCoMo. Dhravya notes that while current benchmarks are useful, they often fail to measure true personalization and the nuances of episodic, user-specific information.
22:26 - Hybrid Memory and Cost Optimization: To ensure high performance, Supermemory has introduced a "hybrid mode" that combines structured memory retrieval with RAG fallbacks. Dhravya emphasizes that optimizing infrastructure for cost is critical for developer adoption, noting that they have achieved costs as low as 2 cents per million tokens.
25:05 - Future Outlook: Looking ahead to 2026, the team is focusing on prosumer-grade interoperability between different AI providers, launching plugins for tools like Cursor, and expanding into voice agent partnerships.
github.com/supermemoryai/openclaw-supermemory
https://x.com/DhravyaShah/status/2023630749065228364
Timestamps
0:00 - The Evolution of Supermemory: Dhravya explains how his project started as a consumer bookmarking/note-taking app and evolved into a core infrastructure layer for LLM memory, focusing on user profiles, temporal reasoning, and knowledge graphs that avoid the inefficiencies of traditional triplets.
8:26 - The Problem with Current Memory Approaches: Dhravya critiques how OpenClaw and similar tools use static Markdown files and agentic tools for retrieval, which often leads to missing context, lack of freshness, and inconsistent performance. He proposes a hooks-based approach that dynamically injects relevant context.
16:32 - Benchmarks and Evaluation: The discussion covers the state of memory benchmarks like LongMemEval and LoCoMo. Dhravya notes that while current benchmarks are useful, they often fail to measure true personalization and the nuances of episodic, user-specific information.
22:26 - Hybrid Memory and Cost Optimization: To ensure high performance, Supermemory has introduced a "hybrid mode" that combines structured memory retrieval with RAG fallbacks. Dhravya emphasizes that optimizing infrastructure for cost is critical for developer adoption, noting that they have achieved costs as low as 2 cents per million tokens.
25:05 - Future Outlook: Looking ahead to 2026, the team is focusing on prosumer-grade interoperability between different AI providers, launching plugins for tools like Cursor, and expanding into voice agent partnerships.










