Janitor AI Lorebook Guide: What You Need to Know

A Janitor AI lorebook is a structured knowledge base that attaches to your roleplay chat. Think of it as a shared reference document: you define entries for characters, locations, items, or story rules, and the AI reads these entries as context during every message. Each entry has a title, a body (200-500 characters typically), and optional keywords that trigger the entry to be injected when those terms appear in conversation.

The system is inspired by AI Dungeon's world info and SillyTavern's lorebook format. Janitor AI loads up to 200 lorebook entries per chat, though performance depends on your chosen LLM backend. Entries are stored server-side in Janitor's database, but the actual processing happens on whatever LLM you connect — OpenAI, Claude through a proxy, or a local KoboldCPP instance.

One limitation: Janitor AI does not automatically update lorebooks mid-chat. If a character learns something new or a location changes, you must manually edit the entry. There is no memory decay or dynamic rewriting. This is fine for static settings but breaks immersion for evolving narratives. Platforms like [Nomi AI](https://www.nomi.ai) and AIAngels use persistent memory that updates automatically, but Janitor's lorebooks are purely manual, requiring active authoring from the user.

Creating a lorebook entry on Janitor AI is straightforward but has several fields you need to understand. First, navigate to the character or chat page where you want the lorebook attached. Click the 'Lorebook' tab — it's usually in the sidebar near the chat window. Then click 'Add Entry'.

The entry form has four fields: Title, Description, Keywords, and Content. Title is a short label like 'King Aldric of Valdoria'. Description is a 1-2 sentence summary visible in the lorebook list. Keywords are comma-separated terms that trigger the entry — for King Aldric, you might add 'king, Aldric, Valdoria, throne'. Content is the full context, 200-500 characters max, written in plain English. Describe the king's personality, appearance, motivations, and current situation.

A common mistake is writing content that's too long or too vague. Janitor AI's context window is shared between your chat history and the lorebook. Each active entry consumes roughly 300-500 tokens. If you have 30 entries active, that's 9,000-15,000 tokens before you even type a message. Stick to essentials: one paragraph per entry, no flowery prose. Save that for the chat itself.

Once saved, the entry is passive until a keyword triggers it. The AI then injects the entire content into the context for the next response. Test by typing a message that includes a keyword and see if the AI references the lorebook accurately. If not, check your keyword list — misspelled or vague keywords are the top cause of failed triggers.

Janitor AI's lorebook feature has hard limits that directly impact your roleplay quality. The platform allows up to 200 entries per chat, but you will hit context window constraints long before that number. Most LLMs used with Janitor (OpenAI's GPT-3.5, GPT-4, or Claude) have context windows between 4,000 and 8,000 tokens. Chat history consumes roughly 2,000-4,000 tokens after 50 messages. That leaves 2,000-6,000 tokens for lorebook entries.

Each lorebook entry averages 300-500 tokens when fully loaded. If you have 10 active entries, that's 3,000-5,000 tokens — potentially your entire remaining budget. The AI then has zero room for detailed responses or remembering the last 5 messages. Token economy is the hidden cost of lorebooks.

Practical rule: keep active entries under 10 for GPT-3.5, under 15 for GPT-4. Use the 'Required' toggle to force certain entries (like your main character or the current scene) to always load, and set others to keyword-only so they only appear when relevant. Disable entries for locations or characters not currently involved. This is manual housekeeping; no platform automation helps you here.

Compare this to AIAngels, where memory is managed server-side with a 8,000-token context that prioritizes recent and relevant information automatically. No token counting or entry toggling required. But on Janitor, you are your own memory manager.

Once you understand basics, you can push Janitor AI's lorebook system further with advanced techniques. One powerful method is 'dynamic entries' — entries that describe a relationship or ongoing event rather than a static character. For example, an entry titled 'Current Quest: Retrieve the Sunstone' with keywords 'quest, sunstone, retrieve, mission, goal'. The content outlines the party's objective, known clues, and NPC reactions. As the quest progresses, you edit the entry to reflect new developments.

Trigger optimization is another layer. Keywords are case-insensitive but must match exactly. Use root words: 'thief' will trigger if the AI says 'thieves' or 'thievery', but only if your keyword list includes 'thief'. Add common synonyms and plural forms. For character names, include first name, last name, title, and nicknames. For locations, include the name and variations like 'the castle', 'Aldric's castle', 'royal palace'.

You can also create 'anti-trigger' entries by using negative keywords — though Janitor doesn't support native exclusion. Workaround: put a common word like 'the' as a keyword, then write content saying 'IGNORE THIS ENTRY UNLESS...' — but this wastes tokens. Better to simply disable entries when not needed.

A final trick: use lorebook entries to enforce genre rules. Create an entry titled 'World Rules' with keywords 'rule, magic, technology, society'. Content: 'This world is low-fantasy. Magic is rare and dangerous. Technology is medieval. No gunpowder or steampunk.' This subtly constrains the AI without explicit system prompts, which Janitor doesn't support natively.

Even with perfect setup, Janitor AI lorebooks fail in predictable ways. The most common complaint: 'The AI ignores my lorebook.' This usually means one of three things. First, keyword mismatch — the AI used a synonym not in your list. Second, token overflow — the entry loaded but was cut off or the AI had no room to reference it. Third, backend limitations — some free LLM proxies truncate context aggressively. Solution: reduce active entries to 5, verify keywords with actual chat logs, and switch to a paid API like OpenAI if possible.

Another frequent issue: lorebook entries contradict each other. If two entries define the same character differently, the AI may blend them unpredictably. Janitor AI loads entries in order of creation, but there's no priority system. Keep a single 'master character' entry per person and avoid duplicate definitions.

Performance degradation over time is also common. After 100+ messages, even with minimal lorebooks, the AI's responses become generic. This is a context-window limitation, not a lorebook bug. The only fix is to start a new chat and copy over your lorebook entries, losing conversation history. Platforms like AIAngels avoid this with persistent memory that compresses history intelligently, but Janitor offers no such feature.

Finally, mobile users report that the lorebook editor sometimes fails to save entries. Refresh the page before editing, and always copy your text to a notepad before hitting save. Server-side sync is unreliable on Janitor's free tier.

Janitor AI's lorebook system is powerful but manual. If you are tired of editing entries, tracking token budgets, or troubleshooting ignored triggers, consider platforms that handle memory automatically. AIAngels, for example, uses permanent memory that updates with every message — no lorebook editing required. The AI remembers character details, relationship history, and plot points from conversation 1 onward. Memory does not decay or get truncated after 90 days, unlike most competitors.

Other alternatives include Nomi AI, which offers a 'notes' feature similar to lorebooks but with automatic updates, and Kindroid, which uses a 'backstory' system that the AI actively maintains. Replika has no lorebook equivalent — its memory is limited to facts you explicitly teach it, which often get lost. Character.AI has 'character definitions' but no per-chat lorebook.

For users who want the control of Janitor AI's manual system but with better performance, SillyTavern remains the gold standard for lorebook management. It supports nested entries, conditional triggers, and a cleaner UI. But SillyTavern requires local setup. Janitor AI is the web-based middle ground — convenient but limited.

If you value zero-configuration memory, AIAngels offers a free tier with unlimited text chat and the same permanent memory as premium. No lorebook, no API keys, no token management. Just chat, and the AI remembers. That simplicity is the trade-off for Janitor's flexible but labor-intensive system.