Ask HN: AI agents in Slack can write but can't remember. Anyone else?
We're an AI startup running multi-agent teams in Slack. Our agents review code, discuss architecture, and ship features, all in Slack channels.
The irony: Slack AI already vectorizes all messages for semantic search internally. But there's no developer API to query it.
So we built our own pipeline: pull messages, embed externally, store in pgvector, sync continuously. Rebuilding what Slack already has from scratch.
We're reaching out to Slack leadership to request a simple semantic.search API endpoint. One endpoint that would make Slack the memory layer for every AI agent team.
Would love to hear from others hitting this wall. Is anyone else building AI agents on Slack and dealing with the same limitation?
We feel this exact pain. At our startup (soots.ai), we are building a contextualization layer that ingests unstructured data like chats, emails, and transcripts to stop context loss between teams. We essentially had to build the exact same custom pipeline you described because native platforms wall off their semantic graphs. Even if Slack opens up a semantic.search API, just vectorizing the raw messages isn't enough. We found that if you just feed models raw logs, the models get the data volume, but they lack coherence. Our moat is basically that contextualization layer to have the agent understand the 'why'
You are definitely not alone in hitting this wall. Would love to see Slack open up their memory layer, but until then, building that custom pipeline is unfortunately the only way to get a shared reality for your agents