Vector memory
Han AI can store a document — a contract, a report, a long email thread — and recall it later by meaning, not by keyword. The store is a local Chroma instance running in Docker on your VPS.
What it does
Embeds a document with OpenAI’s text-embedding-3-small and writes it to a Chroma collection. Later turns can query by natural-language intent and get the most relevant chunks back.
| Field | Value |
|---|---|
| Schema names | remember_document, recall_documents |
| Powered by | Chroma in Docker (v2 API) |
| Embedding model | OpenAI text-embedding-3-small |
| Data location | Your VPS, in the Docker volume |
When Han AI uses it
- A document is too long to keep in every turn’s context and you still want it referenced later.
- A pattern of past conversations should be searchable (“the supplier we talked about three weeks ago”).
- A reference library — playbooks, term sheets, supplier files — needs to be available on demand.
Examples
- “Remember this MSA — I’ll ask questions about it later.”
- “Pull anything we have on the Phnom Penh renovation project.”
- “What did we decide about payment milestones last quarter?”
Limits
- Embedding calls cost tokens (charged against your monthly budget at the embedding rate).
- The chunks Han AI sees are the top matches, not the whole document — for full-document review, use document extract on the original file.
- Smoke test is pending on production tenants.
TODO: confirm vector memory E2E verification on first content-heavy tenant.
Why this stack
Chroma is the leading open-source vector database, runs in a single Docker container, and keeps your embeddings on your own disk instead of shipping them to a hosted vector service.