Sources

Memcal reads the streams you already get and archives new items into its local, append-only store. Start with one source, verify it with a small ingest, then add the rest.

Source	Page	Credential
iMessage	Setup	BlueBubbles password, or local Messages database
WhatsApp	Setup	None — local macOS app database
GroupMe	Setup	API v3 personal token in ~/.memcal/.env
Slack	Setup	User OAuth token (xoxp-…); memcal login slack provisions it, or paste by hand
Telegram	Setup	telegram_api_id / telegram_api_hash plus one interactive login
Signal	Setup	memcal login signal (QR scan) or signal-cli link -n memcal
Email	Setup	Proton Bridge credentials in ~/.memcal/.env
Calendar	Setup	macOS permission via memcal ical setup
Agent conversations	Setup	None — flows from the integrations

Credentials live in ~/.memcal/.env (one key=value per line) — you never edit it by hand for these: each memcal login pastes, validates, and saves for you, asking before it overwrites anything. memcal sources reports what each source still needs; memcal doctor checks the whole setup.

Comparison

	Slack	Telegram	Signal
Auth type	User OAuth token (xoxp-…); memcal login slack provisions and validates	api_id/api_hash + one interactive memcal login telegram (phone → code → 2FA), session file after	Linked device via memcal login signal (QR scan on the phone)
Library	slack_sdk>=3.27 (pip install "memcal[slack]")	telethon>=1.36 (pip install "memcal[telegram]"); pip install "memcal[chat]" gets both	No pip package — signal-cli JVM binary (brew install signal-cli), optional signal_cli= path override
Incremental model	PolledSource: per-conversation watermarks (slack.<channel_id>), newest-first history paging	PolledSource: per-chat watermarks (telegram.<peer>:<id>), newest-id shortcut skips quiet chats with no request	StreamSource: single destructive drain, one high-water cursor (signal.received) that cannot rewind
History availability	Full channel history, oldest-first per round (page 200, bounded by --limit/--rounds); pre-first-run dormant chats (> 30 days silent) skipped until they speak	Full dialog history via min_id forward walk; broadcast channels excluded; dormant chats (> 30 days) skipped on first run	None before the link. Only the undelivered queue from the link moment forward; no backfill exists
Rate limits	SDK retries (3× rate-limit, 2× connection) honoring Retry-After; 429 stops the round with rate limited by slack — stopping this round, watermarks committed	FloodWaitError stops the round with rate limited by telegram — stopping this round; memcal never retries through the limit	No API rate-limit handling; the bounds are receive --timeout 10 and a 180s run ceiling, surfaced as signal-cli timed out after 180s
Risks	Token sees all DMs incl. private channels — guard .env; re-runs are safe and idempotent	Session file is full account access — guard ~/.memcal; re-runs are safe and idempotent	Destructive queue: a crash between receive and commit loses messages permanently; --limit does not slice a drain; two homes must never share one link

All three share the ingest economics: --limit (default 1000 items/round), --rounds (default 25) via catch_up, read / archived / queued summaries with [more waiting], N passed but older than 30d spool-horizon notes, and N skipped as muted for platform-muted threads (muting shapes priority and presentation, never collection).

Which order to set up

1. Slack first if you have it. No interactive login, fastest verify loop (memcal sources → memcal ingest slack --limit 20), and the richest history.
2. Telegram second. Needs the my.telegram.org app page and one interactive login; session reuse makes every later run non-interactive. Good history depth once logged in.
3. Signal last, and promptly. No history exists before the link, and the queue starts accumulating at link time — link, verify (linked as …), and ingest in one sitting so nothing waits acked-but-unarchived.

After each source: memcal ingest <name> --limit 20, read the summary, then memcal ingest all and memcal dream --dry-run before adding the next one. If a source misbehaves, its page's Troubleshooting section lists the exact error strings the code produces; Troubleshooting covers the cross-source checklist.