A Proactive AI Companion
for Sustaining Long-Term Goals

Our Mission

Reinforcement learning for human improvement

The Problem

The Intention–Action Gap

92%

Failure Rate

According to a 2024 Forbes Health survey, 92% of people who set long-term goals never achieve them — that's 9 out of 10 people. The problem isn't willpower or laziness. It's the enormous distance between wanting something and doing something about it, day after day, month after month.

Three forces work against you

Four Pillars of Intentive

Data Science Techniques

Our Data Science Approach

Multi-Goal Intent Mapping

One of Intentive's core innovations is reducing the cognitive friction of daily tracking. When you write a single check-in — something as casual as “I ran 3 miles and listened to a Python podcast on the way” — our LLM-based classification layer automatically identifies every goal that check-in is relevant to. The user simply writes a short update and the system takes care of the rest.

Benefit 1

Maximizes signal extraction per interaction from minimal user input

Benefit 2

Reduces cognitive load by eliminating redundant goal-by-goal tracking

Dynamic Goal State Detection & Re-engagement

To determine when intervention is needed, Intentive tracks each goal using a dynamic goal state model with adaptive stall detection. If a user does not engage within the inactivity window, the goal automatically transitions to a stalled state, triggering re-engagement mechanisms such as nudges and context-aware suggestions.

State	Description
Active	Default state when goal is created. Intentive will try to return user to this state.
Paused	User has chosen to pause. Intentive recognises this and will not nudge.
Stalled	Inactivity detected. Triggers nudges, context suggestions, and local event recommendations.
Completed	User marks complete. Goal achievement plan is finished.
Archived	User archives from any state. Goal achievement plan is finished.

Research-Driven Interventions

Intentive uses a RAG pipeline built on a behavioral science research corpus (PubMed, ArXiv — 5 papers, 278 chunks) to surface scientifically proven goal-achievement tips. This separates Intentive from apps that only provide generic encouragement or re-hash popular beliefs.

Other apps say…

“Believe in yourself! Keep trying and have confidence you can do it! Get sleep and eat vegetables. Eye of the Tiger!”

Intentive says…

“Create Specific Implementation Plans. Develop clear, detailed strategies for overcoming potential challenges in training. Specify exact times, locations, and actions.”

Architecture

System Architecture & Privacy

Intentive is built on a 5-layer stack with clear separation of concerns. The architecture is designed around one core idea: you stay in control — always.

Privacy, Security & Guardrails

Privacy wasn't an afterthought — it shaped the architecture. The separation between Memzero and the main backend is a privacy decision: memory lives in its own isolated service with its own auth, so access is explicit and auditable. The right to be forgotten maps to a real delete endpoint.

Content Guardrails

Five harm categories are screened at ~0ms latency via regex pattern matching — no LLM call required. Every pattern and action is documented in GUARDRAILS.md.

Agentic Layer

Agents

Intentive has 15 agents across 4 categories, 12 of which are LLM-powered. The key design decision is the two-tier approach — LLM-powered agents handle work that requires reasoning, while rule-based agents handle deterministic tasks faster and cheaper. All agents communicate results via SSE for real-time UI updates.

Memory Layer

Long-Term Memory

Most apps store your data in rows and columns — and we do that too with PostgreSQL. But the interesting layer is Qdrant, our vector store. When you write a check-in, we convert it into a 768-dimensional vector embedding and store it alongside your goals and past memories. Later, when you write another check-in, we retrieve the semantically closest past context and give it to the AI as background. This is how the assistant gets smarter over time — it's not just reading your latest message, it's reading it in the context of everything relevant you've ever written.

Evaluation

Evaluation Strategies

The core question: would a real user stay on track with their goals because of this system?

We built a two-layer evaluation framework to measure whether outputs are useful, personalized, and well-timed. All results are tracked in Braintrust for comparison across prompt iterations.

Layer 1 — Agent Prompt Evaluation

12 eval modules test each agent prompt in isolation. Heuristic scorers check format and calibration; LLM-as-judge scorers measure empathy, actionability, and specificity. The Prompt Engineering Lab (built in Streamlit) enables rapid iteration: select a prompt, run it against Bedrock or Ollama, and compare outputs side-by-side.

Heuristic scorers

Format compliance, calibration, length

LLM-as-judge scorers

Empathy, actionability, specificity, persona-fit

Layer 2 — Simulated 30-Day User Journeys

We defined 7 synthetic personas, each with distinct behavior patterns. A LangGraph agent runs each journey day by day — an LLM role-plays as the persona and decides whether to check in based on what the system sent them. Nine real agent prompts fire in sequence.

Fading Felix

Loses motivation over time

Busy Blake

Always exhausted, inconsistent check-ins

Comeback Casey

Two-week break, then returns strong

+ 4 others

Various engagement and dropout patterns

7 scoring dimensions

CoherencePersona-fitStrategy calibrationGoal progressFailure modes (6 checks)RetentionNudge reactions

Combined threshold: 0.7+ = system is working for this persona.

Key Learnings & Impact

Key Learnings & Impact

Acknowledgements

Acknowledgements

Future Work

From Prototype to Product