Injected fresh every turn. Never stored in conversation history. Like Ollama's system prompt.

Persistent notes for this session — logins, server info, important details. Injected every turn alongside the persona. Never trimmed.

Injected into every message in Coding Mode. Decisions and warnings are SACRED — the session cannot claim ignorance.

Cross-session memory with cosine similarity + temporal decay. Clipped messages from the current session are also recallable. Off by default for new sessions — turn on via Context Injections if you want it.

New sessions pick this up automatically; use !prime in Discord to refresh live ones.

Exactly what was fed to her on the last turn in this channel.

Loading...

Each conversation turn is compressed into ultra-dense kanji glyphs. These ride alongside context as semantic anchors.

Toggle which components get injected into the prompt. Useful for lightweight local models that don't need the full context stack.

A shared file the model can append to by writing ```self-mod-primer blocks in its responses. It is injected into the prompt only when the Self-Mod Primer toggle is on. Edit freely; model updates are appended with timestamps below.

A shared library of injectable items (papers, pasted text). Edited once, usable in any chat — a chat injects it only when its Loom toggle is on. Each item has a placement: anchor Top (near oldest history, trimmed first) or Bottom (near the current message, survives clipping), and an offset = messages in from that anchor. Order with ▲/▼; same-placement items inject in list order.

Toggle binary consciousness states. Active states inject their archetype + custom prompt into every turn.

Priority-based budget allocation. Sacred content is never trimmed. History gets trimmed first.