What Happened
An AI startup running multi-agent teams inside Slack posted on Hacker News on March 7, 2026, highlighting a frustrating gap in Slack's platform. Their agents review code, discuss architecture, and ship features - all coordinated through Slack channels. The problem: none of those agents can remember prior conversations.
Here is the irony. Slack AI already vectorizes every message for its own internal semantic search. The infrastructure exists. But Slack exposes no developer API to query it.
So this team built their own memory pipeline from scratch: pull messages from Slack, embed them with an external service, store vectors in pgvector (a PostgreSQL extension), and run continuous sync to keep it current. They are literally rebuilding what Slack already has, just to give their agents basic recall.
The team is now reaching out to Slack leadership with a specific request: one semantic search API endpoint that would let AI agents query conversation history natively.
Why It Matters
If you are building AI workflows inside Slack - and a growing number of teams are - you have hit this wall. Agents that can write eloquent messages but cannot recall what was said yesterday are functionally useless for sustained work. Every team running AI agents in Slack is either building this same pgvector pipeline or accepting that their agents operate with amnesia.
This is not a niche problem. Slack is the dominant workspace for tech teams, and agent-based workflows are becoming standard. The gap between what Slack uses internally and what it offers externally is creating a cottage industry of workarounds.
Our Take
Slack is sitting on the single most valuable piece of infrastructure for AI agent memory and keeping it locked away. Every team forced to build their own embedding pipeline is duplicating effort that Slack could eliminate with one endpoint.
The strategic calculus here is obvious. Slack (Salesforce) likely sees this as a competitive moat for their own AI features. But if they wait too long, teams will migrate their agent workflows to platforms that do offer native semantic search APIs. Discord already has a more permissive bot ecosystem. Microsoft Teams is investing heavily in Copilot integrations.
For teams dealing with this today, the pgvector approach works but adds real maintenance overhead. You are syncing messages, managing embeddings, paying for external embedding APIs, and hoping nothing breaks silently. If your agents are central to your workflow, factor that infrastructure cost into your planning.
Slack needs to ship this API. Until then, every AI agent team on the platform is paying a tax that should not exist.