PRDY — Persistent Reflective Dialogue of Yours

The name

PRDY stands for Persistent Reflective Dialogue of Yours. The name encodes the design intent: the dialogue is persistent (memory accumulates), reflective (always WHY, never WHAT), and yours (this is not a generic assistant — it is shaped entirely by the person using it).

It wasn't designed to feel like a product. It was designed to feel like a thoughtful friend who listens well, asks the right question, and remembers what you've told them. That aspiration shaped every design decision: one question per response, always why not what, memory that accumulates across sessions, no summaries of what you said back at you — just genuine engagement with the next layer of the thing you're exploring.

The design

PRDY is structured as 11 chapters — each one exploring a different dimension of life, career, and thinking. Within each chapter, the interaction follows one strict rule: one question per response. And a second rule that shapes what the questions are: always WHY, never WHAT.

What did you do is a reporting question. It produces a list. Why did you do it is a reflection question. It produces understanding. PRDY isn't building a résumé. It's building a picture of how I think, what I value, what I've learned, and how I've changed.

Not "what did you work on at The Minimalist?" but "why did building systems at scale matter to you at that point in your life?"

Memory across sessions

The system uses prady_memories — a table in the icuboid Supabase project. At the start of each session, 4–10 relevant memory chunks are retrieved and injected into context. This means PRDY knows what was discussed two months ago. It can follow up. It can notice patterns across time.

This is a design decision with real consequences. Most conversational AI has no memory. Each conversation starts from zero. PRDY's design premise is that the relationship — the accumulation of shared context — is the value. Without memory, there is no relationship. Without a relationship, the WHY questions stay surface-level.

Phase 2: semantic memory

The current memory system is keyword-based retrieval. Phase 2 activates when the memory store reaches 50+ entries: pgvector similarity search replaces keyword lookup, enabling semantic rather than literal memory recall. The question "what makes you feel most alive at work?" can retrieve memories about The Minimalist, about the Vipassana retreat, about the IVR project — because the system understands conceptual proximity, not just string matching.

Phase 2 also connects PRDY to the portfolio layer. The memories accumulated through reflective dialogue become the substrate for the Ask Rohan agent — a way for the portfolio itself to answer questions about who I am and how I think, grounded in actual reflective work rather than a written bio.

What it's investigating

PRDY is research into three intersecting questions:

The first is about memory architecture. What memory retrieval strategies produce the most coherent long-term conversational relationships? Keyword, semantic, hierarchical, chronological — these produce different kinds of conversation. The right architecture for a reflective companion is not obvious and can only be evaluated through sustained use.

The second is about WHY-based interviewing. The hypothesis is that consistently asking why rather than what, over time, produces a different kind of self-understanding than most reflection tools. Testing this requires the tool to actually be used over time — which is what PRDY does.

The third question is about AI companionship and identity. Can an AI system that accumulates memory about a person help them understand themselves better than they could alone? Better than journalling? Better than occasional conversations with thoughtful humans? These are real questions that require real use to investigate.

What it uses

PRDY runs on claude-sonnet-4-5 — chosen because Sonnet produces the best balance of conversational warmth, WHY-question depth, and cost efficiency for a system that runs across many sessions. The interface is a simple password-gated web app at apps.icuboid.in/prady.html. The password keeps it personal — this is not a public product.

n8n orchestrates the conversation: incoming message → memory retrieval from Supabase → Claude Sonnet with memory context + chapter prompt → response → memory write for significant new material. The loop is simple. The value is in the accumulation.