Add Slack as a memory source

[ellistarn]

Context

Current memory sources (OpenCode, Claude Code, Kiro) are all local agent conversations with clear user/assistant structure. The reflect pipeline extracts judgment patterns from human corrections to AI — the conversation structure gives us signal for free.

Slack is fundamentally different:

• API-based, not local files
• Multi-party, not user/assistant pairs
• Continuous streams, not discrete sessions
• Low signal-to-noise — chatter mixed with judgment

But Slack captures something the agent sources don't: how you think in your own voice when you're not correcting an AI. Design decisions, architectural arguments, pushback on bad ideas — that's in Slack threads, but it's not structured as corrections.

Plan

1. Introduce a Source/Provider interface

The current code hardcodes each source as a function call in Upload() (muse.go:157-176). Before adding Slack, refactor to a proper interface:

type Provider interface {
    Name() string
    Sessions() ([]Session, error)
}

Upload() iterates over []Provider instead of source-specific if blocks. This is overdue regardless of Slack — we already have 3 sources following the same pattern.

2. Slack source adapter

• Use Slack API (bot token or user token) to pull threads where the user has substantive participation
• Thread = session (threads have enough coherence to be a unit of memory)
• Filter out noise: reactions, one-liners, link drops
• Map thread participants to roles, preserving who said what
• Env var for token: MUSE_SLACK_TOKEN, user ID: MUSE_SLACK_USER

3. Adapt the dream pipeline for non-agent sources

The current reflect pipeline assumes human→AI correction patterns. Slack needs a different extraction model:

• What you contributed to a thread matters, not the correction pattern
• May need a different reflect prompt or a parallel extraction path keyed on source
• The learn phase (soul synthesis) can likely stay unified — reflections are reflections regardless of where they came from

This is the hard part and may need experimentation with prompts before it's clear what works.

Open Questions

• Should Slack extraction be a separate pipeline stage, or can we adapt the existing reflect prompts to handle multi-party conversation?
• What's the right unit of memory — threads only, or also DMs and time-windowed channel activity?
• Rate limits and incremental sync — how to avoid re-pulling everything on each push
• Token scoping — what's the minimal Slack permission set needed?

[jamesmt-aws]

on the multi-party conversation topic, we should talk about the importance of identity across working sessions. really like this one, this morning I was thinking about Github PRs as another source of memories

[ellistarn]

I consider gh to be truly poisoned by ai content at this point :)