So instead of markdown files in a repo, it's markdown files in a nosql database?

Its not Markdown files in nosql. The context is extrapolated from a session and broken out into a record Postgres. You decide in a session what memory/context to save to the vault, the next time you ask about it, you will receive that (and refresh the AIs current context in the process).

Consider the non-developer type jobs out there who may be using AI, but the AI their team is using has no clue what the other person is doing or has done. This tool helps them bridge that gap. If one person has a pile of markdown files (not talking about skills, agents, etc) and the others do not, they won't get the same benefit. Instead of having hundreds of 'solutions' or 'learnings' in a file, you store it in the cloud and depending on their role/group assignment, its available to their AI client to recall later.

▲

infogulch 19 hours ago | parent [-]

I'm not trying to be obtuse, and correct me if I'm wrong, but "context" is just words, often rendered as markdown because it's easy for both LLMs and humans to produce and edit. Since its in a database and they're not necessarily full documents, I'll use the phrase "markdown snippets" for now. Feel free to quibble with the phrasing, I'm trying to understand what it does.

You save 2 respectability points by using postgres over nosql. (joke)

> depending on their role/group assignment, its available to their AI client to recall

So you generate/extract markdown snippets from the local context of each agent along a few axes like: what problem the user is working on and their approach, what role they're acting as and tools they're using, etc, and store those snippets in a postgres db with indexes over the category.

So how is that then exposed to the receiving agents? Is it fully opt-in / pull only? Or do they get some related threads automatically injected in their context. What does it look like? "Jan was working on something similar yesterday, her agent found these docs helpful: a.md, b.md" ?

▲

Repeater22746 18 hours ago | parent [-]

Yes, sorry, context is such an overused word these days. What happens is the conversation is distilled into several areas (problem, solution, learnings, 'context' or original problem, plus other fields) and vectorized. Later, when you ask something similar, it will search and rank what it finds as potentially most relevant. This helps others discover what you or others already had done or learned. I leveraged aspects of Postgres that help rank, index and boost recall in a way having plain markdown files can't exactly achieve.

> what problem the user is working on and their approach, what role they're acting as and tools they're using, etc, and store those snippets in a postgres db with indexes over the category.

30,000 foot level - yes.

The receiving agents connect to the MCP, and when you say things like 'save this to the vault', 'save to your memory', or 'did someone solve this already?' it will use the exposed MCP tools and store the information. The AI agent is responsible for distilling the conversation bits out to the storage shape, then inserted to Postgres. This way, its AI client agnostic, there is no AI client lock-in, and teams that use a mixture of clients can retain the service all the same and experiment/deprecate AI clients as they go without losing any data.

Its kind of hard to present real world examples, but I figured this is very useful for even non-developers. Like your local plumber or HVAC company or a hundred other examples, "Customer in 14A electrical panel shorted again - what did we do last time to resolve that?" - record is found and presented back in conversational form from AI for that user to work through, if that makes sense. I tried to imagine professions "in the field" out on a call and thought wouldn't it be cool if they could get answers to things faster than the hours it can often take to diagnose issues. That is why it was important for me to support the desktop/mobile app versions and not just CLI clients to reach those kinds of customers using a UI.

Also yes, if it finds say 6 potentially relevant records, it will rank them and present them back as options to explore - it won't force #1 on you just because it was ranked #1.

Basically, I put myself in the shoes of an end user, potentially not as tech savvy as the HN crowd, who do not want to create or maintain markdown files (they may not even know what it is) and particularly do not use CLI. They can still use this tool with desktop/mobile apps, not worry about the overhead of managing several files (not to mention sharing them to others), and the more they use it the better it becomes.

I hope this helps answer your questions.

	▲	infogulch 15 hours ago \| parent [-]
		So you draw some distillates out of the conversation, do a vector embedding of each, and insert the pairs into pg. Then retrieval is distill -> find top N nearby vectors. Indexing, etc. Seems like a smart way to organize long term memory. Also I can't help but notice that this solution combines technologies whose origins literally span 6 decades: relational databases (80's), vector embeddings (00's), and LLMs (20's). What kinds of prompts do you use to get the distillates, and what shape are they? I guess I've been assuming you extract more than one distillation out of a conversation per memory access, do you?