The basic shape is to periodically "distill the conversation into several areas (problem, solution, learnings, 'context' or original problem, plus other fields) and vectorized" (aka vector embedding), then queried against pgvector table to find related "memories". The vectorized distillates are also inserted into the pgvector table with a reference to back to the source conversation to add new memories.

Vector search requires a full scan but it's still pretty fast and I bet it's more accurate the FTS.

I don't see a downside yet.

Eventually the obvious spam callers would all quit and the volume would reduce to a handful of borderline cases. So the spam department wouldn't make a lot of money anymore, but the volume would be so low it would barely register.

For example say you get a spam call. You notify your service provider and pay $10 to have them review the call recording to verify that its spam. An employee listens to the call and determines that it is, in fact, spam. You get a $100 reward for reporting spam and your service provider forwards the charge (+10%=$110) to the peer network where the spam call originated. They will in turn forward it to the next peer, or if the originator is a customer, charge and/or close the offending account.

If this system is mandatory then the problem will quickly sort itself out.

I fixed this and a couple related issues in a PR but it hasn't gotten any attention yet. I guess they are a bit swamped. https://github.com/zed-industries/zed/pull/59937

Some changes I made at the direction of an agent: fix domainkeys CNAMEs for M365, rotate M365 dkim keys that haven't been rotated for over a decade, fix broken spf record formatting.

My biggest issue is that squarespace refuses to enable dkim signing for transactional emails that they send for us (order/shipping confirmation etc). The email sending service they use (socketlabs) supports it but they are not interested in enabling the feature, so I can't lock down our dmarc. I guess that means squarespace is not a good fit for our needs; it's just disappointing that we have to move to a different platform again for technical reasons that are solvable with a dashboard switch.

Totally normal gardener behavior.

As for "subdue", ask anyone maintaining a large garden how much they have to work to keep it from reverting back to a wilderness. Gardening is choosing which species stay and which are forcibly removed and optimizing the environment to enable your chosen species to thrive. If the words "subdue" and "dominion" freak you out, pick your brains up off the floor and go touch grass.

Western civilization has decided that it wants to keep the primate species and is dissatisfied when they become roadkill, so it altered the environment to enable the chosen species to thrive. This article is a gardening success story. Gardening requires dominion. We have dominion over the primates and we have chosen to be benevolent with that dominion.

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?

This makes me wonder if the reason why agents love weird punctuation is because the labs run the base models through a RL training step that forces them to correct their grammar; but instead of rewriting short spliced sentences into long coherent sentences, they just learn to splice them together with punctuation that passes the automatic grammar checker.

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" ?

Now LLMs come along and they also have their own phrasing preferences. But now it's a problem because what used to be personal preferences of a single person that manifests in 5000 words per day from one person tops, is now the bias of a single model multiplied x10,000,000,000 generated tokens per day so any bias sticks out like a sore thumb.