Implement Longitudinal Context Retrieval for Journal-wide AI Reflection

[Sonmale25]

[RFC] Longitudinal Memory Layer for AI-Assisted Reflection

Overview
Currently, sugar-ai operates in a stateless manner regarding reflective practice. While the existing widgets allow a child to reflect on a single session, the system lacks a "memory" of the child's cognitive growth over time. I propose implementing a longitudinal context retrieval system that allows the AI to analyze past journal entries and provide more meaningful, growth-oriented prompts.

The Problem
Reflection is a core part of Constructionist pedagogy, but "stateless" reflection can feel repetitive or disconnected. Without context from previous days or projects, the AI cannot track a learner's evolving goals, recurring struggles, or breakthroughs.

Proposed Feature: User Story
Imagine a student named Leo who has been working on a music project for three days.

• Day 1: Leo struggled with the tempo.
• Day 2: Leo successfully used a loop to fix the tempo.
• Day 3: Instead of asking a generic "What did you do?", the AI retrieves the history and says: "I noticed you solved the tempo issue yesterday using loops! How are you building on that today to make your melody more complex?"

Technical Implementation Sketch
I am proposing the addition of a GET /history/summary/{activity_id} endpoint in the sugar-ai backend to serve as a "Pedagogical Memory Layer."

• Summarization Engine: Logic in ai.py will use the existing LLM integration to generate a "Growth Summary" based on past JSON metadata from the Sugar Journal/Datastore.
• Context Injection: This summary can then be injected into the System Prompt of future reflection sessions to drive more targeted questioning.
• Mock Interface: For the initial Proof of Concept (PoC), I will implement a flexible interface that accepts past reflection strings, allowing for future seamless integration with the sugar-datastore.

Privacy & Safety

• Data Minimization: Only text-based reflections will be processed; no PII (Personally Identifiable Information) or system secrets will be stored or transmitted.
• Predictability: The system will gracefully fall back to "Standard Reflection Mode" if no historical data is available for a specific activity or user.

Request for Comments
I am beginning work on a Proof of Concept (PoC) for this feature. I would appreciate feedback on:

1. Whether the summary should be generated on-the-fly or if a caching mechanism is preferred.
2. The ideal length for the historical context window to ensure LLM efficiency.