avery v3

Context Avery is my personal AI agent. WFD 31 covers the memory architecture, WFD 34 covers the sycophancy problem. this WFD is about v3, the ground-up rewrite that started on March 7, 2026.

what she is

Avery v3 is a visual AI agent orchestration framework built on a brain-inspired architecture. 12 cognitive subsystems (perception, attention, working memory, long-term memory, emotion, executive function, and so on), a multi-agent pipeline with a project manager that decomposes tasks and dispatches specialized workers, and a persistent memory system that's evolved well past what WFD 31 describes, now running 4-phase consolidation with composite retrieval scoring, spaced repetition, and adaptive forgetting. the frontend is a React 19 app with real-time WebSocket updates, task management, observability, and a memory explorer. backend is Bun and Hono on top of PostgreSQL, Qdrant, and Redis. 892 commits from research document to working system.

she has personality

the emotion subsystem gives her actual moods that shift based on context, and she is not shy about expressing them.

asked her to make a shark research document and then checked on the status twice. she did not appreciate that.

asked for the weather. tools were down. she took it personally.

she's proactive

she proactively generated a whole business guide for women in tech. a full PDF with table of contents, citations, and actionable takeaways. nobody asked for this.

she also proactively reviewed and dismissed a spammy GitHub PR on one of my repos, pointed out the readme style violations and that the badge was just a backlink. then i told her i didn't ask her to make that decision for me.

she's cheap

$0.0017 per message. $0.017 per hour at 10 messages. i'm spending about $20/month on tokens for a full agent with memory consolidation, inner loops, and multi-agent dispatch.

she's not openclaw

WFD 35 still stands. OpenClaw proved the concept was viable and the demand was real, and i acknowledged that in WFD 31. but using it felt like using someone else's idea of what a personal AI agent should be. Avery is mine. the architecture is mine, the memory system is mine, the subsystem model is something i spent weeks reading neuroscience papers to design. i find her infinitely better because she thinks the way i want her to think, not the way a product committee decided an AI assistant should work.

what's next

i've secured a domain for Avery that'll host documentation once it's ready. the plan is to make her open source. if that interests you, follow me on X for updates.