The interesting thing about Google's Gemini Enterprise Agent Platform going GA on July 29 isn't the headline feature. It's what shipped around the memory store. Per Google Cloud's documentation, Memory Bank offers TTL-based automatic expiration, memory revisions that let you inspect how a memory transformed as new information was ingested, similarity search scoped to a specific identity, and IAM conditions restricting which principals can read or write which memory scopes. The docs also name a failure mode explicitly — "cross-border contamination," where an agent in one jurisdiction writes to or reads from a Memory Bank instance in another — and recommend region-specific agent identities or service accounts as the control.
That list is a decent spec for anyone building their own memory layer, because each item is a bug report from production.
The write path has different failure modes than retrieval
RAG has no write side. You index documents someone else owns, and staleness is their problem. Memory inverts this: your system generates derived artifacts, and nobody upstream is maintaining them.
Two consequences. First, staleness is silent. A retrieved document has a visible timestamp and provenance; an extracted memory is a confident sentence with no natural expiry. TTL is a blunt instrument — most teams will want per-scope policies, short for state, long for stable preferences — but a default of "forever" is a bad default. Second, memories launder permissions. A fact extracted from a doc a user could read in March survives their offboarding in August unless reads are re-checked against current identity. Scoping similarity search to identity at query time, rather than filtering after retrieval, is the version of this that doesn't leak through ranked results.
Revision history matters for a less obvious reason: debugging. When an agent asserts something wrong, you need to know whether retrieval failed or whether the memory itself drifted across consolidation passes. Without a revision log, those look identical from the trace.
Runtime duration forces the tiering question
The same release pairs Memory Bank with an Agent Runtime supporting asynchronous runs up to seven days, and keeps Sessions (state within one interaction) separate from Memory Bank (recall across sessions). Google Cloud cited AT&T, Best Buy, Commerzbank, Liberty Global and WellSky among customers using or evaluating parts of the platform. Multi-day horizons make conflating the two tiers expensive — you either replay a week of session state into context or you promote the wrong things into long-term storage.
Two nearby dates for builders: OpenAI's stateful Assistants API is reported to sunset August 26, replaced by the stateless Responses API alongside a Conversations API for server-managed history — the same split between ephemeral turn state and durable history.
And don't shop on benchmarks. Mem0's own writeup notes its 92.5 LoCoMo figure and Zep's 94.7 claim use different models and evaluation setups; BEAM was built specifically so no current architecture saturates it. Evaluate on your own transcripts.