This seems way too complicated and unnecessary. Agents are perfectly capable of discovering memories on the FS, following agent instructions.
I guess this adds indexing and querying but most coding agents have good solutions for this already, and it works automagically for everything, not just memories.
What we could use instead is a file system layout standard, which could subsume memories and a lot more. I don't think that's needed either, but it would probably solve more problems than this.
I would love to know how many countless others on HN, like me, find themselves reading about a very they have built and have been using for months talked about like it’s a revolutionary new idea.
IIUC, the most basic version is when you have a log where every entry has both “date added” and “effective date,” so you can add stuff to the log retroactively. For example, “the user just informed us yesterday that they moved last year” -> address date added=yesterday, date effective=last year
I have similar setup in Orgzly (kinda in Emacs too but it's buggy and not not as useful there) where a note has a "created time" property that's always automatically applied. And then there's the "closed time" applied when I set note the state to "done", which I sometimes modify depending on what the note is for and thus what "done" means.
Average people build their own harnesses, and imagine themselves the pioneers of industry. They propose protocols. They code, feverishly, into the night, driven by their vision for the future.
It used to be that 'idea guys' were limited by execution. We now feel the avalanche of these ideas, even maybe executed half-decently, fall upon deaf ears and zero market.
Saw this shared a few days ago, skimmed it, didn't understand it. See it again now, another skim, still don't understand. I think it could use a ELI15 or something.
Sorry to be a debbie downer, but this reads like LLM slop rather than engineering work. I don't just mean the language on the page-- although that too (not an X it's a Y, over and over again)-- but the absence of the artifacts of ActualEngineering(tm) rather than just a flood of vibes.
For example, I would expect to see tables or figures showing task success rates on some benchmarks for agents augmented with and without this proposal, perhaps before and after fine tuning, or running against alternatives or to the extent that there are no alternatives against variations of this design that were considered and rejected.
Otherwise what reason is there to think that this design is better than some alternative or even any good at all? Perhaps it causes agents to hallucinate like crazy-- who knows if it hasn't been tested.
Work like that is what makes efforts like this worth sharing and worth reading about-- anyone can spend a few minutes and ask their favorite LLM to design such a framework and get something that looks "credible". But in a post LLM world credible alone is externally indistinguishable from anti-social time wasting slop.
AI agents can already use tools and coordinate, but their memory is fragmented across project files, agent notes, local stores, databases, and vendor-specific systems. Move to a new tool and the context is gone.
UMP v0.1 is a shared format plus a simple way to read, write, update, and move agent memory across tools. The goal is memory that's user-owned, auditable, and extensible across agents and runtimes, instead of locked inside one vendor.
It's early (v0.1) and I'd love feedback on the format and where it breaks down. Repo and spec are linked from the site.
I guess this adds indexing and querying but most coding agents have good solutions for this already, and it works automagically for everything, not just memories.
What we could use instead is a file system layout standard, which could subsume memories and a lot more. I don't think that's needed either, but it would probably solve more problems than this.
with apologies to Andy Warhol - in the future, everyone will have a universal protocol for agent memory that is on the HN front page for 15 minutes.
0: https://github.com/edihasaj/universal-memory-protocol/commit...
MCP came from Anthropic, A2A from Google so they had big tech backing from day 1.
As a developer, I wouldn’t touch this without confidence I can get gains down the line from interoperability.
What is a bi-temporal record? I don’t think I’ve heard the term before and I’d love to learn more.
Average people build their own harnesses, and imagine themselves the pioneers of industry. They propose protocols. They code, feverishly, into the night, driven by their vision for the future.
It used to be that 'idea guys' were limited by execution. We now feel the avalanche of these ideas, even maybe executed half-decently, fall upon deaf ears and zero market.
https://github.com/space-bacon/SRT
I can read any models every thought. No one cares. Not the narrative.
For example, I would expect to see tables or figures showing task success rates on some benchmarks for agents augmented with and without this proposal, perhaps before and after fine tuning, or running against alternatives or to the extent that there are no alternatives against variations of this design that were considered and rejected.
Otherwise what reason is there to think that this design is better than some alternative or even any good at all? Perhaps it causes agents to hallucinate like crazy-- who knows if it hasn't been tested.
Work like that is what makes efforts like this worth sharing and worth reading about-- anyone can spend a few minutes and ask their favorite LLM to design such a framework and get something that looks "credible". But in a post LLM world credible alone is externally indistinguishable from anti-social time wasting slop.
> Memory is attacker-controllable input. The spec requires a verify, filter, frame rehydration pipeline. Never string-interpolated into the prompt.
Uhhh... so who wants to tell them how LLMs work?
UMP v0.1 is a shared format plus a simple way to read, write, update, and move agent memory across tools. The goal is memory that's user-owned, auditable, and extensible across agents and runtimes, instead of locked inside one vendor.
It's early (v0.1) and I'd love feedback on the format and where it breaks down. Repo and spec are linked from the site.
any other feature being compatible between harnesses makes transitioning from one to another too easy
so, the only way memory will work, similar to {AGENTS,CLAUDE}.md, is if everyone uses: base path + markdown files