Memory and traces

The backend keeps three things about players: a session, a lasting memory, and traces of runs. Each lives only as long as it needs to, and each belongs to one server: the same player on another server is someone else (see Security and privacy).

#The session

Each player has a session in Redis, kept 24 hours after the last question (CHAT_TTL_SECONDS).

It holds whole exchanges: the question, the tool calls with their results, and the answer. A follow-up like “how many were in that chest?” needs no second call.
Once a session passes about 32 000 tokens or 100 exchanges, everything but the newest 12 000 tokens (at least 2 exchanges, at most 30) folds into a summary of up to 3000 characters. The fold runs after the answer, so the player doesn't wait.
Only a session past 42 000 tokens folds before the answer.
A player who left gets no summary: the session simply expires.
Between folds the session only grows at the end, so the provider's prompt cache keeps it.
/askx requests land in the session as one line: what was asked and which commands came out.

The v1 and v2 APIs share the session.

#Other threads

Public questions have their own thread, one per server. Each question starts with the asker's name, so a friend's follow-up makes sense. It keeps the answers but not the tool results, which only the asker may see. It lives 2 hours after the last question.
Automation agents each have a thread apart from the player's session, for 2 hours after their last wake.

#The lasting memory

The lasting memory sits in MongoDB until the player clears it, or 365 days after their last question. It holds up to 30 short facts: where the base is, what the player works on, how they like answers. The agent saves them itself and says so.

Players manage it with the plugin's commands:

/allay new starts a fresh session; memories stay.
/allay memory lists the facts; with text, it saves one by hand.
/allay forget deletes them.

An operator gets a copy of everything about a player, or erases it, with /allay privacy.

#Traces

The backend traces runs into its MongoDB through Mastra's observability: each run, each model step (prompt, output, tokens), and each tool call with its arguments and result. Mastra Studio reads them.

Every run is traced by default, since 6.4 on hosted backends too (TRACE_SAMPLE).
Traces go after 7 days (TRACE_DAYS). A trace holds players' questions, so it lives no longer than it needs to.
Keys are redacted. A run takes about 150 KB.
The admin panel shows only a trace's shape: steps, models, tools and whether they worked, tokens, and timings.
Erasing a player deletes the traces of their runs at once.

Next: Troubleshooting