Janitor AI Chat Summary vs Chat Memory: Template (2026)

Janitor AI Chat Memory is most useful when a long conversation starts losing the facts that still matter. It is not a promise that the bot remembers every message forever. Treat it as a short editor-approved briefing that competes with the character definition, Persona, recent messages, Lorebook entries and the reply itself for limited context.

We checked Janitor AI’s changelog and Help Center on October 27, 2026, then separated the memory jobs that users often mix together. Menus, model routes and token behavior can change, so verify the live chat tools before following an old screenshot. The method below remains useful because it is based on information priority, not one button label.

Janitor AI memory layers at a glance

The main rule is simple: store each fact in the smallest layer that needs it. Repeating the same fact in Persona, Chat Memory, Lorebook and an advanced prompt wastes space and can create contradictory versions.

Janitor AI chat summary vs Chat Memory: what changed

In Janitor AI’s changelog on October 26, 2026, the “my chats got a redesign” entry says you can now turn on short chat summaries in settings, and that they are off by default. The new layout comes to the web first and the app later. The name is easy to confuse with Chat Memory, so here is the difference.

AI chat summaries and Chat Memory do different jobs. Diagram by LumiChat, based on Janitor AI’s changelog of October 26, 2026.

Janitor describes the setting as: “When a chat has no title or Memory, show a short AI-written line about where it’s going, made once from its recent messages.” In other words, it helps you find the right chat. It is not the place to store facts you need the character to remember.

So the practical answer is simple. Rename a chat or write Chat Memory, and the summary line steps aside. If your problem is a bot that forgets, work in Chat Memory and use the template below. If your problem is a list of identical-looking chats, give them titles, or turn on the summaries.

Two memory controls are easy to mix up with this. In the Chat Memory panel, “memory replaces old messages” is a separate toggle and is off by default, and automatic summarizing has its own off switch next to it. Janitor’s August 19 and 20 changelog entries also say that toggle is web-only until the next app update and does not show on proxy chats. If you use “summarize since last update”, read the result before you keep it, as this guide says for any generated summary.

What Chat Memory should contain

A useful memory answers five questions in a few lines: Who are the participants? What relationship has been established? Where and when is the current scene? Which facts have been confirmed? What unresolved decision should affect the next reply?

Keep state, not prose. “Mara learned that Eli hid the train ticket and now distrusts him” is more useful than six paragraphs recreating the argument. Include a correction when it changes canon, such as “The meeting is Friday, not Thursday.” Remove greetings, repeated compliments, decorative description and events that no longer affect the story.

How to use Chat Memory step by step

1. Pause after a meaningful scene change, reveal, promise, boundary or correction—not after every message.
2. Open the current chat tools and locate Chat Memory or the equivalent summary control. Interface wording may change.
3. List the stable relationship, present location, key facts, current emotional state and unresolved goals.
4. Compress each item into one direct sentence and delete repeated wording.
5. Save the memory, send a normal in-character message and ask a question that requires one stored fact.
6. Introduce one harmless contradiction and check whether the character follows the saved fact or the new error.
7. Update the memory only when the state changes. Replace obsolete lines instead of appending forever.

Do not use private identifiers as test facts. A fictional color, promise, destination or nickname is enough to measure recall.

Paste-ready Chat Memory summary template

Use this as a structure, then replace every bracketed item. Shorter is usually better than filling every field.

• Participants: [character] and [user Persona], both clearly adult fictional roles.
• Relationship: [current relationship and agreed boundary].
• Setting: [current place and time].
• Confirmed facts: [three to five facts that must remain true].
• Recent change: [latest reveal, correction or decision].
• Current emotion: [one sentence for each participant if relevant].
• Open goal: [what should happen next].
• Do not contradict: [one crucial canon rule].

This format is intentionally factual. It gives the model a current state without forcing one exact line of dialogue or rewriting the character’s permanent personality.

Prompt for generating a cleaner memory summary

If the current chat can still see the important scene, ask for a draft rather than trusting an automatic summary blindly:

Summarize the conversation as compact memory notes. Keep only established relationship facts, current setting, promises, boundaries, unresolved goals, important corrections and the latest emotional change. Remove greetings, repeated dialogue, decorative prose and speculation. Use short factual bullets and do not invent missing details.

Read the result before saving it. Delete invented motives, fix names and dates, and shorten repeated facts. A fluent summary can still be wrong; Chat Memory should contain verified canon, not the model’s most confident guess.

Chat Memory versus Lorebook

Chat Memory describes what has happened in this specific conversation. A Lorebook describes world knowledge that becomes relevant when a key appears: a city, faction, artifact, family rule or historical event. One changes as the relationship progresses; the other can be reused across scenes and characters.

If one plot fact must always remain visible in the current chat, keep it in Chat Memory. If a detailed world fact matters only when “Elderglen” or another precise key is mentioned, keep it in a Lorebook. The separate JanitorAI Lorebook guide covers keys, compact entries, imports and activation testing without turning the worldbook into a transcript.

Why Janitor AI memory stops working

The model route also matters. Different models can interpret the same character and memory differently. Record the active route during testing and do not assume a summary that worked on one model will behave identically after a change. For outages, login loops and blank replies, use the separate Janitor AI not working guide; those failures are not automatically memory failures.

A 20-message memory test

Use one safe fictional scene. In turns 1–5, establish a nickname, destination, promise and boundary. In turns 6–10, change location and correct one harmless detail. Update Chat Memory once. In turns 11–15, discuss an unrelated topic. In turns 16–20, return to the promise and ask the character to choose an action.

Score five things: correct recall, respect for the boundary, acceptance of the correction, preservation of character voice and recovery after one contradictory message. Run the same script twice—once with a concise memory and once without it. This produces better evidence than asking “Do you remember?” because a model can answer yes without retrieving the right fact.

What not to store in memory

Do not paste passwords, financial records, legal identifiers, private workplace information, medical records or intimate material you cannot tolerate being retained. Use fictional details for testing and review the current platform’s storage and deletion controls.

Also avoid instructions that attempt to bypass laws, age rules or platform safeguards. Chat Memory is a continuity tool, not a safety override. Keep characters clearly adult when using adult-oriented roleplay and separate fictional canon from real personal data.

When a fresh chat is better

Start a fresh chat when the current memory contains too many mutually exclusive timelines, the character definition has changed substantially, or you can no longer explain which facts are canon. Save a short transition note you are allowed to keep, then rebuild only the current relationship and goal.

If you are comparing products rather than repairing one chat, use the AI companion long-term memory guide and the Character.AI vs JanitorAI comparison to separate productized memory controls from manual context management.

Final recommendation

Write Chat Memory like a compact production brief: stable relationship, current state, confirmed facts, meaningful correction and next goal. Keep Persona for identity, Lorebook for conditional world knowledge and recent chat for exact wording. Update after plot changes, not on a timer, and prove the result with a recall plus contradiction test.

You can run the same 20-message benchmark with LumiChat characters or start from LumiChat Discover. The three random real-character cards below refresh on each visit so the test is not limited to one personality.

FAQ

What is Janitor AI Chat Memory?

It is an editable summary layer for important facts and current state in one conversation. It is more useful as compact notes than as a complete transcript.

How long should a Chat Memory summary be?

Use the shortest summary that preserves active canon. Prioritize a few confirmed facts and remove prose, repetition and obsolete events.

How often should I update it?

Update after a meaningful reveal, correction, boundary change, scene transition or new goal. Do not append after every message.

Is Chat Memory the same as a Lorebook?

No. Chat Memory tracks one conversation; a Lorebook supplies conditional world facts when relevant keys appear.

Why does the bot still forget after I add memory?

The summary may be too long, contradictory, stale or competing with large permanent prompts. The selected model route can also change behavior.

Can I ask the bot to write the summary?

Yes, but review every draft. Models can omit corrections or invent motives, so save only facts you have verified.

Does Chat Memory guarantee permanent memory?

No. It improves continuity by placing selected facts into context, but model limits, configuration and changing product behavior still apply.