# Troubleshooting chat > Fixes for empty replies, repetition, forgotten details and rate limits. Source: https://reverie.im/en/docs/chatting/troubleshooting Language: en When a reply fails or comes back wrong, Reverie usually says why in a strip above the message box, with the action that fixes it — **Retry**, **Switch to** another model, **Use recommended model**, **Start new chat** or **New session with memory**. This page covers what each symptom means. **The reply came back empty** The strip reads "Model returned an empty reply — try again or switch models". You are not billed for it: a reply of one token or less is recorded but never charged. Most often the model declined the scene rather than failing. A related message, "Model declined to respond — try rephrasing or switching", means the same thing explicitly. The free model is the most likely to do this — "The free model is small and sometimes declines scenes it can't handle" — and Reverie offers to regenerate the reply on a stronger model. What to do, in order: 1. **Retry** or regenerate. The previous attempt is kept behind the variant switcher, so nothing is lost. 2. **Switch models** from the button in the composer, then regenerate. 3. Edit your last message and rephrase the part the model balked at. **The reply stopped mid-sentence** Two different things look the same. **Reverie cut it off.** If a reply starts looping — repeating a phrase or a paragraph it has already written — the stream is stopped at that point and trails off with an ellipsis. Regenerating usually produces a clean reply; see the repetition entry below if it keeps happening. **The model stopped early.** Use **Continue** on the character's last message. That resumes the reply from where it stopped instead of rewriting it. If the strip says "This chat is too long for the current model — switch models or start a new chat", the conversation no longer fits: switch to a model with a larger context window, or use **Chat Settings → Conversation Setup → Start New Session with Memory**, which compresses the conversation into a background and continues in a fresh chat. **The character keeps repeating itself** Reverie already watches for this in one-on-one chats and adds a correction to the next turn when recent replies echo your wording, open the same way, or reuse the same paragraph shape. In group chats it does not, because some similarity between speakers is expected. When that is not enough: * **One-turn slash commands.** `/Pacing: Fast` moves the scene on, `/Response length: Short` breaks a wall-of-text habit, and a **Custom** directive can name the specific tic. * **Edit → AI Rewrite** on the offending message. Its one-tap actions include **Avoid repeating sentence patterns**, **Avoid repeating paragraph structure** and **Add more variety in expression**. * **Regenerate** for a different take — the old reply stays available behind the variant switcher. * **Switch models.** Repetition is strongly model-dependent. * **Start New Session with Memory** if the whole conversation has settled into a groove. The new chat keeps the memory and drops the pattern. **The character forgot something I told it** Work out which layer lost it. Only the most recent 10,000 tokens — and at most 50 messages — stay in the conversation verbatim. Everything older survives as a summary, or as a saved memory, or not at all. 1. **Check the Memory tab.** If the fact is not there, add it, and **Pin** it. Pinned entries are sent in full on every turn and cannot be rewritten or deleted by the AI. 2. **Check Conversation Memory** for plot details rather than facts. You can edit a summary directly, or **Rebuild** it. 3. **Check that memory is on.** **Learn from Chat** is a per-character switch; when it is off the character neither saves nor reads memories. 4. **Check the model.** On a model without tool support, a character cannot save memories mid-conversation. Saved memories are still used. 5. **On a free account** there is no automatic background pass — mid-reply saving and the panel are the only routes. Bounty-challenge chats have memory disabled entirely. If you see "Memory limit reached", that scope is full at 100 entries; edit or delete some before adding more. See [How memory works](https://reverie.im/en/docs/chatting/memory). **You've hit the free model limit** The free model allows **15 messages every 3 hours**, counted per account and shared across web chats, stories, novels, the bots and the API — switching surfaces does not give you a fresh allowance. Subscribers get a larger share of the same limit: three times as many on Pro, four on Premium, five on Ultimate. The count itself is shared, so upgrading mid-window raises the ceiling but does not reset what you have already used. Your options are to wait out the window, switch to a paid model, or subscribe. See [Free models](https://reverie.im/en/docs/reference/free-models). **Not enough credits** The credit check runs *before* a reply starts, so a reply already streaming finishes; it is the next message that is blocked with **Not enough credits**. From there you can **Switch to free model**, top up, or subscribe. Two things worth knowing while you decide: an empty reply is never charged, and the credit multiplier of every model is shown on the model button before you send. See [Credits](https://reverie.im/en/docs/billing/credits). **Reply suggestions stopped appearing** When enabled, a batch appears after a new character reply finishes streaming. Free suggestions are an introduction, not a daily allowance. Your first 5 batches are free in total, whether you have a subscription or not. The allowance never resets. Guests also receive a one-time allowance for their browser identity; signing in uses the account’s separate allowance. Each batch is one generation of three suggestions. Later account batches use credits. After the final free batch, automatic suggestions turn off once. An automatic paid batch leaves about one reply’s worth of credits on the default model untouched. If there are too few credits, automatic suggestions pause until a top-up or page reload; there is no daily reset. The notice offers **Get credits**. You can still request suggestions manually with the wand. New accounts and first-time guests start with automatic suggestions on. Accounts that existed before this change start with them off; old browser settings are not imported. One switch covers all characters. Account settings sync across devices; guests keep their setting in this browser. **Side panel → Settings → Chat Experience → Chat Reply Suggestions** [Reply suggestions](https://reverie.im/en/docs/chatting/smart-suggestions#what-they-cost) **A model is failing or shows an Issue badge** An **Issue** badge next to a model in the picker means there is an open incident on it: "This model currently has an open issue. Check the status page for details." A degraded or outage incident also raises a banner inside the app, and an outage banner stays until the incident is resolved. The error strip tells you which kind of failure you hit — "Model is busy right now" (with a countdown before **Retry** re-enables), "Network hiccup — couldn't reach the model", "Model is having trouble responding", or "Model took too long to respond". During an incident the fastest fix is **Switch to** another model or **Use recommended model**; your conversation is unaffected. `/status` lists every component, any active incident and the incident history. ## Related \[#related] * [How memory works](https://reverie.im/en/docs/chatting/memory) — pinning, summaries and why the live window is small * [Choosing a model](https://reverie.im/en/docs/reference/choosing-a-model) — which model to switch to, and why * [Credits](https://reverie.im/en/docs/billing/credits) — what a reply costs and how to top up * [Getting help](https://reverie.im/en/docs/account/getting-help) — when none of this is it * [Why AI gives empty responses](https://reverie.im/en/blog/why-ai-empty-responses) and [Breaking the repetition curse](https://reverie.im/en/blog/breaking-repetition-curse) — the design notes