workflow memory

workflow memory is our research into how agents should remember the way work actually gets done.

the problem

most memory systems either remember too much or too little. too much memory becomes noise. too little memory makes the agent repeat the same questions, rediscover the same constraints, and miss the habits that make a user productive.

useful memory should not be a transcript dump. it should be a compact model of preferences, recurring decisions, project conventions, and known pitfalls.

what belongs in memory

• stable project conventions
• repeated user preferences
• operational constraints
• decisions that should not be relitigated
• recurring commands and workflows
• mistakes the agent should not repeat

what should not belong

• temporary chat context
• secrets
• stale implementation details
• speculation
• anything the user would not expect to persist

where this is going

we are exploring memory as a versioned workspace artifact rather than a hidden model feature. users should be able to inspect it, edit it, delete it, and understand why the agent used it.