Shared agent memory stopped being a per-session feature this week. Databricks and CollectivIQ launched it, on consecutive days, as a shared layer for other agents and employees to use.

Databricks unveiled Agent Memory on 22 October, in beta in its Agent Bricks workspace, according to TechTarget. CollectivIQ rolled out GigaBrain on 21 October, a shared intelligence layer that connects knowledge its AI teammates gain across projects, processes and customer relationships, according to its launch announcement.

The direction is the same in both cases. Memory no longer belongs to one agent in one session. It accumulates, and the next agent starts from it.

Shared agent memory moves next to the data

Databricks built Agent Memory on its MemCurrent architecture and runs it on Lakebase, its serverless PostgreSQL database, so that, in product manager Michael Bendersky's words reported by TechTarget, "customers keep their memories in their own environment."

The problem it answers is repetition. Without memory, Bendersky said, agents repeated mistakes, forgot user preferences and needed to rediscover facts with each session. Analyst Sanjeev Mohan of SanjMo put it more plainly to TechTarget: "Models suffer from amnesia. Even a stateless agent can show up to work every morning with no recollection of what it did yesterday."

Databricks reported, in testing that TechTarget noted was not independently verified, that the design helped agents complete three times as many enterprise tasks and reduced token consumption by approximately one-third.

The scale context matters. TechTarget cited McKinsey's State of AI in 2026 report showing that just 40 percent of organizations with revenue over $1 billion are now able to scale agents, and only 22 percent of those under $1 billion are successfully building agents. Memory is being sold as one reason that gap exists.

Databricks paired the memory launch with AI Runtime, generally available, with governance through Unity Catalog. Analyst Kevin Petrie of BARC U.S. told TechTarget that the larger problem remains elsewhere: enterprises "have to solve the bigger problem of validating, preparing and governing the context that feeds those models."

What GigaBrain adds, and what it leaves open

CollectivIQ starts from the handoff, not the database. Its AI teammates, which it calls Digital Direct Reports, develop context as they handle assignments. GigaBrain makes relevant parts of that context available to other authorized employees and agents, so a second teammate does not reconstruct a customer's procurement process from scratch, according to Unite.AI's review of the launch.

The launch material includes controls that shared memory needs. The system can trace stored information to its source and applies access permissions when sharing it. An incognito mode lets employees keep sensitive interactions out of shared memory. Its product page, as Unite.AI reported, says facts carry their source and who confirmed them, single-source claims await corroboration, and stale facts expire.

Unite.AI also noted the limits in the public material. CollectivIQ published no quantified benchmark for accuracy or productivity, and it did not specify the corroboration rules or expiry intervals. Those details decide whether a shared fact is a reviewed record or a summary that other agents repeat.

That distinction applies to derived knowledge as well as source files. Withholding a document is not enough if its contents reappear in an accessible summary. Permissions have to travel with what memory creates, not only with what it read.

The governance gap is measured, not guessed

The market for shared memory is arriving before most organizations govern context well. In Redis's State of Context Engineering report, 94 percent of organizations surveyed said compounding system intelligence is essential for production-grade maturity. Only 4 percent said they had reached the stage where that actually happens.

The same report found that 69 percent of the most mature organizations still feed critical decisions with stale data, 60 percent said they are ineffective at governing who and what can reach context, and only 58 percent were confident their context is well governed.

Shared memory amplifies both sides of that record. A correct, current fact reused by ten agents saves ten rediscoveries. A wrong preference, an expired price, or a one-off exception stored as a permanent rule is also reused ten times, faster and with more confidence than a person would repeat it.

This is the point yesterday's pillar on context engineering made for working context, and it applies with more force to a shared record. Working context can be rebuilt for the next task. A shared record persists. It needs provenance on every fact, a date when the fact stopped being true, a correction that propagates to agents that already used it, and a review step before a draft becomes the record other agents trust. The scale problem in the Volkswagen deployment was the same problem at fleet size: counts of agents do not settle whether what they share is true.

The launches this week are therefore best read as infrastructure catching up to a record-keeping problem. Putting memory next to governed data, as Databricks does with Lakebase and Unity Catalog, and tracing facts to a source and a confirmer, as CollectivIQ describes, are the right starting points. Neither settles who reviews a new fact before it joins the shared record, how long it stays, or how a reader opens it and checks it without asking an agent.

That is the work a company brain has to do. Shared agent memory is useful only when the memory is a record people can open, read and correct, with the same permissions and provenance as the sources it came from. The agents will reuse whatever is in that record. The question for buyers is no longer whether their agents will share memory. It is who keeps the shared record true.