My coding assistant forgets everything I told it the moment a session ends, and the ChatGPT tab next to it never knew any of it. Fixing that for myself costs about $52 a month on AWS in eu-west-1 for
Markdown supported · ⌘↵ to post
The vendor-independent part feels like the real win here. I've built enough small AI tools to know how annoying it gets when useful context is stuck in one place. Keeping it portable makes a lot of sense.
Really interesting approach to solving the memory fragmentation problem across AI tools. I like that the article is upfront about the limitations, especially the lack of SSO, high availability, and independent security audits.
The per-person credentials and spend caps also make this more practical for team adoption. One question: how do you plan to handle memory quality over time as multiple agents update shared knowledge and potentially introduce conflicting or outdated information?
The "where does this memory live" question is the one that decides everything downstream. I hit the personal-scale version of it with a local document tool: the moment the honest answer is "on someone else's server", half the documents you actually need help with are off the table. Your $52/month in your own AWS account is the team-scale version of the same answer. Curious about the off-switch: when you revoke one person's access on a Friday afternoon, does their stored context stay readable by the team, or is it scoped per user?
Really interesting approach to solving a problem that’s becoming more noticeable as teams start using multiple AI agents and different model providers. I especially liked the idea of keeping the memory in the company’s own AWS environment rather than locking it to a single vendor.
The combination of shared memory, version history, access controls, and the ability to switch between agents makes this feel quite practical for real-world engineering teams. Also appreciated the transparency around what Memrain doesn’t support yet, especially SSO, HA, and compliance.
The one-week pilot approach is a nice touch too — testing the actual workflow, costs, permissions, and recovery process before committing feels like the right way to evaluate something like this.
BomToBe
BOMTOBE makes orke, a desktop workspace for Claude Code and Codex.
The clarification about retrieved passages versus cited answers is useful: source provenance alone does not prove that an answer is supported. For an eval set, I’d score both exact page/version attribution and claim-to-passage support, including stale or conflicting pages and mixed exact/semantic matches. Then compare each answer with the pointers the agent received. Disclosure: I work on Orke at BOMTOBE, a desktop workspace for Claude Code and Codex.