HGPT

What it is for

A self-hosted AI workspace where conversations, projects, decisions, and explicit memory remain available even when the selected model changes.

AI conversations are usually trapped inside one provider and important context disappears into old threads. HGPT keeps the durable parts in a database you control, then supplies relevant context to whichever supported runtime you choose.

Keep a long-running project

Carry project instructions, confirmed facts, and decisions across conversations and model changes.

Inspect an answer

Review the recalled context and provider trace rather than treating memory as an unexplained black box.

Try without registering

Start an isolated guest workspace and select an available runtime.

Get started

Open the live workspace, create a conversation, and select an available model. Add a project when you want instructions and decisions to persist beyond one chat.

• A modern browser and a selected available model runtime.
• For self-hosting: Node.js, PostgreSQL, and the repository deployment configuration.
• An API-backed runtime needs credentials; a local runtime needs its model installed.

Features and workflow

One workspace, many runtimes

Use local Ollama, installed Codex or Claude CLIs, or a configured OpenAI-compatible API without moving the conversation.

Durable memory

Save confirmed facts explicitly and connect them to projects, topics, conversations, and decisions.

Inspectable recall

See which memories, earlier messages, topics, and provider were used for an answer.

Project and decision records

Give long-running work persistent instructions and record choices with rationale, evidence, alternatives, and status.

Private guest access

A signed random guest identity creates an isolated workspace without requiring an account.

Provider controls

Add compatible providers, validate their model catalogs, and keep saved API keys encrypted.

1. Open HGPT and start an isolated guest workspace.
2. Choose a local, CLI, or API-backed model.
3. Create a project or conversation and add durable instructions when needed.
4. Save only the facts and decisions that should persist.
5. HGPT retrieves relevant context, sends it to the selected runtime, and records the trace with the response.

Architecture

The interface is vanilla HTML, CSS, and JavaScript. Express assembles context and streams responses; PostgreSQL is the authority for conversations and memory; local embeddings and topic relationships support recall; Redis serves only ephemeral HAI state. Host-authenticated CLI runtimes are reached through a Unix-domain socket rather than copying their credentials into the application container.

Privacy and limitations

Every query is scoped to its signed-in or guest owner. Guest cookies are HTTP-only and SameSite=Strict. Saved provider keys use AES-256-GCM encryption, compatible provider URLs are screened against private-network and credential-bearing targets, and recalled text is labeled as untrusted context rather than executable instruction.

Status

Live on myhai.org and under active development. HGPT is a personal knowledge workspace, not an autonomous truth engine: the user decides what becomes confirmed memory, and model output still needs normal review.

Availability

Live web app

Best for

Long-running AI work

Data model

User-owned PostgreSQL

Interface

Responsive web workspace

Questions