Back to blog

What breaks when every vendor ships its own agent memory?

Thắng Đoàn
Thắng Đoàn

You want agents that carry context between sessions, so you pick a vendor's memory product. Six vendors shipped six answers this year, and they do not mean the same thing by memory. The risk is incompatibility, not forgetting.

Memory is three different products wearing one name

Same word, three products. Zilliz frames the problem as persistence. Per its release, when a conversation ends most agents forget everything, so memsearch stores long-term memory humans can read and edit. The LavX analysis splits the concept further: context for the immediate task, memory for persistence, knowledge for foundational rules. Google, per VentureBeat, instead treats memory as identity and business context kept across applications for tasks lasting hours or days.

  • Zilliz: human-readable persistence across conversations
  • LavX: context, memory, and knowledge as separate layers
  • Google: memory as identity carried across apps and devices

The definitions collide in practice

Imagine Google-style identity memory meeting Zilliz-style readable files. One agent carries your business context inside a platform-bound workspace. The other exports editable memory files any developer can inspect. Neither can read the other's memory, so cross-vendor workflows still force users to repeat themselves. The fragmentation Zilliz set out to fix reappears one level up.

Where you store it decides who can fix it

Perplexity's Brain stores memory as editable wiki files outside the model, so corrections change behavior without retraining. Unite.AI reports Brain-enabled sessions improved correctness by 9.3 percentage points while using roughly 15% fewer tokens. Entire takes a similar inspectable route. Per ADTmag, its open-source Brain component documents commands for checking saved facts against their sources and inspecting memory health. Memory you cannot read is memory you cannot debug.

Memory has a meter running

The CoReMem preprint from Yale, Tsinghua, and Goodvision AI names the trade most vendors skip. Writing and reading persistent memory costs latency and inference tokens. Teams must weigh the future utility of accumulated context against those costs. LavX makes the same point architecturally, treating context as compiled infrastructure to control token spend.

What this does not solve

Vendors ship storage; you still own the policy. ADTmag notes Entire's Trails evaluation scores require teams to define appropriate approval rules themselves. Unite.AI reports Perplexity's Agent API leaves memory maintenance to developers via versioned Skills and manual context passing. And nothing here settles portability. Per VentureBeat, Google's agent is in private preview and platform-bound, and the sources show no cross-agent memory standard emerging.

If you switched agent vendors tomorrow, which of your accumulated memory could you actually export and reuse?
Who in your team is allowed to correct an agent's stored memory, and how would they even see it?

Write a one-page list of what your agent must remember, who can edit it, and how you would export it.

Sources

ADTmag: Entire Continuous Memory

Unite.AI: Perplexity Brain architecture

LavX: From prompts to context engines

PR Newswire: Zilliz open-sources memsearch

VentureBeat: Google persistent Gemini agents

CASRAI: CoReMem preprint

Share:

Recommended for you

Enjoyed this article?

Subscribe for new articles. No spam. Unsubscribe anytime.

By subscribing you agree to receive the newsletter. No spam, and you can unsubscribe anytime.