Session Gateway is a provenance-first continuity layer for AI agents. It turns separate agent transcripts into one searchable evidence system, while the original transcripts remain the authority.
The system combines five functions:
- It ingests transcripts from different agent harnesses.
- It creates a normalized event catalog with exact source references.
- It distills long sessions into citation-linked task and episode cards.
- It supports lexical, semantic, structured, and time-based retrieval.
- It shows live work through scopes, checkpoints, overlaps, and handoffs.
The result is not a larger chat history. It is a controlled path from a question to a small set of relevant records, and then back to the exact source evidence.
| Document | Purpose |
|---|---|
| Overview | What the gateway does, what it does not do, and why it is valuable. |
| Architecture | The system layers, data flow, and design rules that protect accuracy. |
| Distillation and backfill | How eight concurrent GPT-5.6 Luna agents convert a transcript archive into a semantic research index. |
| Getting started | A practical build order, operating flow, and acceptance checklist. |
native transcripts
|
v
normalized sessions and events -----> live scopes and checkpoints
|
+----> full-text search
|
+----> task and episode cards ----> embeddings ----> vector index
|
question ----> hybrid retrieval ----------------+----> bounded source evidence
|
v
agent reasoning with citations
Three rules keep this design sound:
- Native transcripts are evidence. Cards, embeddings, and vector indexes are derived data that can be rebuilt.
- Models perform semantic judgment. Host code controls identity, bounds, citations, validation, persistence, and concurrency.
- Retrieval is progressive. Search finds candidates; a separate evidence-open step confirms the useful source context.
Start with the overview. Use the architecture guide before implementation.