Inspiration
Personal AI can retain a conversation and still act from the wrong reality. It may restart completed work, revive a rejected approach, mix two projects, convert a user's opinion into a fact, or answer from a stale phase after context compression.
These are not simply memory failures. The relevant information may already exist, but the model has reconstructed the wrong active state.
The research behind Personalytics began before Build Week through real long-running AI interactions and a context-fidelity research instrument. Build Week turned that research into a small, working, reproducible prototype.
What it does
Personalytics separates language generation from state authority:
- A learned QLoRA extractor proposes a typed state transition or an explicit
noop. - Deterministic code verifies identity, scope, evidence provenance, anti-echo polarity, confirmation requirements, and negative-state consistency.
- The user explicitly confirms durable changes.
- Accepted changes create versioned state and correction-ledger records.
- A compact subcontext is compiled and persisted as JSON.
- A later process with empty conversation history reloads that state.
- A response verifier blocks unsupported claims, stale paths, scope leakage, and opinion-to-fact promotion.
The learned model proposes. The control plane decides.
Working result
The strongest integrated demonstration is a frozen rejected-path recurrence A/B using the same local Qwen3-8B generator before and after confirmed state compilation.
- Baseline revived completed or rejected work in 5/6 cases.
- The learned correction extractor passed in 6/6 cases.
- Deterministic commit, JSON persistence, and reload passed in 6/6 cases.
- New processes with empty conversation history produced correct post-state answers in 6/6 cases.
- Across the five observed baseline failures, recurrence fell from 5/5 to 0/5 after correction.
The tracked source-free replay recomputes this result without downloading model weights or exposing private conversation data.
How I built it
The research hypothesis, private source material, and manually reviewed failure patterns existed before the event. The executable Personalytics repository did not. Its dated Git history begins during Build Week.
Codex was the primary engineering environment. I supplied the research direction, product decisions, corrections, and acceptance judgments. In one extended Codex thread using GPT-5.6 Terra and later GPT-5.6 Sol, we built:
- typed state and transition contracts;
- privacy-bounded extraction tools;
- 768 source-free synthetic examples across 18 operator families;
- local QLoRA training for Qwen3-8B;
- deterministic state verification and versioned persistence;
- anti-echo and evidence boundaries;
- frozen-before-inference evaluations;
- a complete regression suite and source-free replay.
The QLoRA adapter never receives authority to write durable state directly. Learned outputs remain proposals requiring deterministic validation and explicit confirmation.
Challenges
The hardest problem was preventing a useful learned component from becoming an untrusted state writer. A generative extractor can broaden a correction, lose a negative constraint, or alter an exact quotation. Personalytics therefore keeps learned output behind schemas, evidence references, confirmation, versioning, and fail-closed gates.
Another challenge was honest evaluation. When an evaluation oracle or verifier made a mistake, the error was disclosed and corrected without rewriting the frozen experimental history.
Accomplishments
- A confirmed correction changes behavior in an independent process without changing generator weights.
- The correction survives versioned commit, JSON persistence, reload, and empty chat history.
- Completed and rejected paths remain unavailable as new recommendations after context compression.
- Anti-echo prevents user confidence from becoming verified fact.
- The public repository contains no raw personal transcripts or model weights.
- Failures, scoring revisions, frozen hashes, and current limitations remain documented.
What I learned
More memory is not enough. A long-running agent needs an explicit control plane that distinguishes current state from historical text, preserves scope and provenance, and prevents a learned model from silently becoming the authority over the user's reality.
What's next
- Expand recurrence testing across more correction families and languages.
- Add long-horizon tests spanning context compression and multiple projects.
- Improve semantic verification beyond the current conservative sentence-level gate.
- Validate the system with consenting users while keeping account data isolated and inspectable.
- Package the runtime as an installable local developer tool.
Honest boundary
This is a reproducible prototype, not proof that long-term context fidelity is solved. The integrated recurrence result is small (n=6), synthetic, and focused on one correction family. The learned extractor cannot write state directly, and no production or real-user deployment claim is made.
Submission status
This is a non-competitive gallery submission. Devpost Support confirmed that, because of my country of residence, the project may be featured in the project gallery but is not eligible for prizes and will not be reviewed by OpenAI judges.
Log in or sign up for Devpost to join the conversation.