A room + short-term conversation-turn memory engine you mount into your own database — not a service, not a pluggable storage backend, just vendor-neutral SQLAlchemy models and the functions that read and write them.
pip install capsize-memoryA Room is any place a bot has a presence: a Discord channel, a Discord DM thread, a Bluesky account's own post stream. A RoomParticipant is one identity seen in a room — a Discord user, a Bluesky handle, or the account itself. A ConversationTurn is one raw utterance, tied to a room and (when known) a participant.
This gives multi-speaker conversations for free: a channel with three humans talking is three distinct participants sharing one room's turn log, not three fragmented one-on-one threads. It also gives a reconstructable log of where a bot has been talking and to whom, across every platform it's wired into — one query, not scattered per-platform ad hoc keys.
from capsize_memory.repository import (
get_or_create_room, get_or_create_participant, record_turn, recent_turns,
)
from capsize_memory.format import format_transcript
room = get_or_create_room(session, "discord", "channel", "guild:channel")
alice = get_or_create_participant(session, room.id, "user123", "alice")
record_turn(session, room.id, "alice", "hey, just moved to Austin", participant_id=alice.id)
record_turn(session, room.id, "capsize", "nice, welcome!")
turns = recent_turns(session, room.id, limit=6)
print(format_transcript(turns))
# - alice: hey, just moved to Austin
# - capsize: nice, welcome!No engine, no connection, no migration of its own — you mount
capsize_memory.models.Base.metadata into your own Alembic env.py
(or create_all() for a quick start) and pass your own Session to
every function. This is deliberate: a second host with its own
database (Postgres, SQLite, whatever) gets the exact same behavior by
pointing its own session at these same models — nothing in this
package assumes SQLite specifically, or owns a connection string.
pip install -e ".[dev]"
pytest
ruff check .
mypy src