Topic / T SALON
AI
T Salon's collection on AI technology — how model capabilities evolve, changing memory system design, observations on AI applications moving from validation to production, and the practical impact on how developers work.
14 stories
- 2026.10.30
MemOS vs. Mem0, Zep, and Letta: Choosing Memory for an Existing AI Product
A comparative look at integrating MemOS, Mem0, Zep, and Letta into existing AI workflows, focusing on direct updates, graph relationships, runtime control, and operational ownership.
- 2026.10.30
Why Customers Repeat Themselves to AI Agents
A destination label does not describe an active support case. How MemOS helps AI agents maintain customer context across handoffs, without losing track of previous troubleshooting steps.
- 2026.10.29
Why AI Agents Need to Know When a Memory Is Out of Date
An agent can recall an old project date accurately and still give the wrong answer. What separates thread memory from stored facts, and how metadata and correction APIs decide which version a later task actually uses.
- 2026.10.15
Agent Memory Observability: 12 Metrics for Writing, Retrieval, Updating, and Forgetting
A practical measurement plan for teams running persistent memory in production, covering 12 metrics across writing, retrieval, updating, and forgetting.
- 2026.10.15
How to Choose a Production AI Memory System in 2026
A decision matrix and evaluation framework for teams comparing open source, self-hosted, and managed AI memory options across lifecycle, retrieval, scope, and operations.
- 2026.10.15
What Grok Build and Claude Projects Reveal About Persistent Agent Memory
An architecture guide for teams that need agents to remember across sessions without turning old context into new mistakes, covering working context, memory lifecycles, and tenant boundaries.
- 2026.10.01
AI Quota Reset Radar: How to Track Codex, Claude & Grok Limits
A practical guide to T Salon's AI Quota Reset Radar (WhenReset). Learn how rolling limits work across Codex, Claude, and Grok, how to project recovery times, and how to pace your weekly budget.
- 2026.09.19
How to Keep Bad Assumptions Out of Agent Memory
Better retrieval alone does not validate a memory that was wrong when written. A research-based look at checking reusable claims, preserving their sources and using memory operations without confusing metadata with verification.
- 2026.09.11
What ChatGPT Dreaming Reveals About Managing Memory Over Years
OpenAI’s Dreaming update puts memory on a months-to-years timescale. We examine freshness, correction and deletion, while distinguishing documented MemOS Cloud operations from modules that remain under development.
- 2026.09.01
Memory Poisoning: How Untrusted Information Persists Across Agent Tasks
Prompt injection can lead to memory poisoning when manipulated content is stored and reused. We distinguish malicious poisoning from accidental contamination and outline six checks for agents that retain information across tasks.
- 2026.08.03
Event Recap: From AI Demo to Production｜Engineering Practices for Agent and AI Native Applications
A record of the August 1, 2026 'From AI Demo to Production' event in Shanghai, with four speakers' presentations and a live demo covering agent memory, model collaboration, Vibe Coding and execution environments.
- 2026.07.21
Cursor’s SQLite Agent-Swarm Experiment: Coordination Lessons and Benchmark Limits
Cursor reports progress on a Rust implementation of SQLite against a held-out sqllogictest suite. We examine its coordination mechanisms and explain why benchmark completion does not establish production readiness.
- 2026.07.20
Hugging Face’s AI-Driven Intrusion Disclosure: Impact, Remediation and Forensic Lessons
Hugging Face disclosed an autonomous agent intrusion into part of its production infrastructure. We separate confirmed impact from open questions and discuss preparation for controlled, locally hosted forensic analysis.
- 2026.07.20
NVIDIA Cosmos 3 Edge: Capabilities, Deployment Requirements and Limits
NVIDIA’s 4B Cosmos 3 Edge brings world understanding, prediction and action generation to physical AI. We examine official deployment conditions, robot integration requirements and device ports that still need validation.