Your agent's memory is now the product. The question is whether it remembers true things

Memory as a feature

Glean built a memory system that carries context across workflows, and the design is deliberately boring in the best sense. Glean uses compaction to summarize conversations into memories, and draws from conversations, work artifacts, and third-party context. Glean uses a process called Mirroring to learn from documents and tickets, and consolidation compares new information with existing memories to merge duplicates. Glean Memory is currently in beta. m:…9259

The system is built to be corrected, not just accumulated. Users can inspect, edit, or delete memories, and saving a memory does not retrain the underlying model. Glean uses a harness to assemble inputs and retrieve memory summaries, and execution traces are used to record agent actions and tool results. Glean Memory uses the Context Protocol (MCP) to interact with other assistants, which means the memory is not locked inside a single product. m:…9259

Glean has filed a provisional patent for its memory system, which tells you how seriously the company takes the idea that memory is a moat. The bet is that the assistant that remembers your preferences, your documents and your tickets is the assistant you keep, and the assistant that starts from scratch every morning is the assistant you replace. m:…9259

Memory that can be wrong

The research side of the week asks the uncomfortable question: what if the memory is wrong. WikiTrace reduces memory-induced degradation by allowing agents to verify lessons against original trajectories, and it outperforms existing memory approaches across five agentic benchmarks. The idea is simple and overdue. A lesson an agent distilled from a past task is linked back to the raw trajectory it came from, so the agent can check the lesson against the evidence when the next task looks similar but is not the same. WikiTrace is described in a preprint, so its results are the authors' own and not yet peer-reviewed. m:…9439

Memory-induced degradation is the failure mode where agents fail because they rely on incorrect or misapplied past lessons, and the evidence-linked approach is designed to catch exactly that. An agent that remembers a shortcut that worked once will apply it everywhere, and the shortcut is usually wrong somewhere. Linking the lesson to the trajectory is the difference between a memory and a superstition. m:…9439

The robot learning world hit the same problem from a different angle this week. Memory-augmented VLA policies often sample frames uniformly across an episode, and stamp each frame with its time step, so time-based memory indexing fails to distinguish between waiting and working. The fix there was to index by events. The fix here is to link by evidence. Both fixes say the same thing: memory is only useful if the agent can tell what it actually remembers. m:…1808 m:…9439

Memory across the enterprise

Databricks is building memory into the enterprise stack too. MemCurrent is a persistent memory architecture for production agents that maintains information consistency, and Genie retains personal memory of user preferences and instructions. Databricks introduced new Genie plugins to provide a governed interface for enterprise context, and the plugins work natively in ChatGPT, Claude, Microsoft Copilot Cowork, and Grok Bot. m:…5218 m:…4911 m:…4914

The plugins carry the governance with them. Genie Ontology combines governed definitions with continuously learned organizational knowledge, and Genie handles data retrieval and governance for AI agents. Agents without native plugins can access Genie via the MCP server, and Genie can route work to domain-specific Genie Agent for specialized analysis. Genie returns interactive visualizations and citations in clients that support the MCP Apps extension. m:…4914

The threads converge on a single idea. Memory is no longer a cache you clear when it gets stale. It is the product: the thing that makes an agent useful on its hundredth task rather than its first, and the thing that makes it dangerous when it remembers wrong. The vendors are selling memory, and the researchers are selling the check on memory. Buy both. m:…9259 m:…9439 m:…5218