Starting a conversation is hard when you don't know what the other side already knows. AI chats have the same problem. You spend the first few messages just setting context, and that's where most conversations lose momentum.
Write Studio's memory system is built to fix this. Three layers, each serving a different purpose.
Layer 1: Who you are (user memory)
Your writing style, your go-to topics, your strengths. This is your writer profile. It barely changes.
Updates: extracted in the background after each Discuss or Suggest session. You never notice it happening.
Injected: into nearly every AI feature at launch. Two exceptions: Discuss only gets it on the first turn, since later turns already carry conversational context. Inspire skips it to stay divergent, drawing from project context rather than user profile.
Layer 2: What you're writing (project memory)
Topic, audience, writing stage, open questions for the current project. Each folder gets its own memory. No cross-contamination between projects. New folders come with a project template. Fill in audience, purpose, and style. The system uses it.
Two update paths. Path A fires after a conversation and captures project state from the discussion. Fast, narrow view. Path B fires on document save and reads everything in the folder and builds a full snapshot.Thorough, but slower. Same format, later write wins. No conflicts.
Injected: into all AI features for the current project.
Layer 3: What you've been up to (cross-project memory)
Overview of all active projects and recent writing. Broader than Layer 2. Updates more frequently than Layer 1.
Updates: after Discuss or Suggest sessions.
Injected: not currently. Tasks like "fix this sentence" or "suggest improvements" don't benefit from knowing what other projects you're running. Adding it would just burn tokens. It will be activate when a real use case appears.
Design principles
- Updates are async and never block the user. They trigger after meaningful actions: finishing a conversation, saving a file.
- Injection is selective. Only relevant layers load. Instructions first, memory second. Total budget: 400 to 500 tokens. Memory informs. It doesn't dictate.
An unsolved problem: memory decay
Updates work by merging old memory with new input. Do this enough times and specifics wash out. The content trends toward generic summaries. Character limits accelerate the problem: the AI preserves broad strokes and drops details when space is tight.
This is a known issue. Future options: re-distill periodically from source conversations, or anchor memory with preserved dialogue fragments.