Most of these words arrived from research papers and never got translated. Knowing six of them turns a chat that mysteriously misbehaves into a problem you can locate.
The two worth learning first are context window and persistent memory, because almost every complaint that starts with the character forgot comes down to confusing them.
The character itself
Persona — the character's identity layer: name, voice, attitude, history, mannerisms. This is what the model is asked to embody on every single reply.
System prompt — the hidden instruction block the model receives before generating anything. Here it holds the persona, the active scene rules and the memory snippet for this specific reader.
Character card — a portable bundle of persona, opening message, default model and sampling parameters. Several community-standard formats exist and are how characters move between platforms.
Memory, and the thing people mistake for it
Context window — the maximum amount of text a model can hold in mind for one reply. Bigger windows reference more of the recent conversation and cost more per call. It is a limit, not a store: when a conversation runs past it, the oldest part simply falls out.
Persistent memory — facts a character keeps about you across sessions. This is a different mechanism entirely: curated, editable, and injected into the context on every reply. Here it is per character, per topic, per reader, and fully visible to you.
This is the distinction that resolves most complaints. A character forgetting something from an hour ago is a context-window effect. A character forgetting your sister's name is a memory-record problem — and if the fact was never written down, no model will recall it correctly.
Token — the atomic unit a model reads and writes, roughly four characters of English. Context windows are measured in tokens, which is why a limit that sounds enormous fills faster than expected.
Steering a conversation
Branch, or fork — a divergence point. Forking from a reply creates an alternative thread that grows alongside the original, and both are kept. This is what makes trying a different direction free rather than destructive.
Regenerate — asking for another version of the last reply without disturbing anything else. Same context, new answer.
The dials
Temperature — how predictable the replies are. Lower is steadier, higher is more surprising. A character that feels flat is often a temperature problem rather than a model problem.
top_p, top_k, min_p — constraints on which candidate next-words the model considers at all. Useful for tuning register: cautious, expressive, terse, verbose.
Repetition penalty — discourages reusing phrases. This is the fix for a character who says their favourite line every fourth reply.
The plumbing
Model router — the infrastructure that picks a model for a scene from a pool and fails over silently when one degrades.
Failover and circuit breaker — the router cutting off a degrading provider before it can poison a conversation with timeouts or refusals, and sending traffic to a healthy one instead. When this works you never notice it, which is the point.
Around the edges
Telegram Stars — Telegram's built-in micropayment unit, bought inside Telegram and spent on upgrades. No card is needed for the top-up itself.
Share link — an attribution-tracked URL that opens a chat with a creator's character. Stars spent by people who arrive through it earn the creator a share.
Topic — a dedicated forum thread in Telegram, used to give each character its own room with its own history and pin strip.
Age gate — the in-bot verification step before mature roleplay can be enabled. The public site is SFW; mature mode is opt-in after the gate, and hard limits stay regardless.
RTL — right-to-left writing, used for Arabic, Hebrew and Persian. Interface elements mirror rather than being left-aligned text in a right-aligned language.