v1

Verifiable Graph

Every Aurra answer ships with a proof tree: the retrieval that ran, the evidence chain, the bi-temporal window, and the supersession history — rendered inline, hash-verified, audit-grade. No more “the agent said so.”

What it solves

Memory layers today return an answer. You don’t see which memories were retrieved, which ones the model actually used, when those memories were valid, or whether they’ve been superseded by newer ones.

That’s fine for a chatbot. It’s not fine for an agent making decisions in regulated environments, working with stale data, or stitching together facts across time. The Verifiable Graph makes the retrieval-to-answer path inspectable on every query.

How it works

Retrieval block

The exact filters that ran — query text, k, similarity threshold, bi-temporal window, source filters — captured before any LLM call. Reproducible. If you run the same query against the same index, you get the same retrieval set.

Evidence chain

Each retrieved memory shows up as an evidence entry with a cited/unused flag, original source citation, decision text, and full provenance. You see which memories the answer actually leaned on — not just which ones surfaced.

Bi-temporal block

Valid-time and transaction-time on every memory. The proof tree records what was true in the world (valid_time) and when Aurra learned it (transaction_time) — separately. Zero date hallucination on LoCoMo, where comparable layers measure 22.95%.

Supersession chain

When a memory is superseded by a newer one, the chain is walked forward and rendered. Cycle-detected, depth-bounded. You see not just the current memory but the history of what it replaced — and what replaced what before that.

Technical details

Endpoint
POST /agent/query
Opt-in flag
include_proof_tree: true
Answer hash
SHA-256, included in proof tree
Proof tree version
1.0
Supersession depth
max_hops = 50, cycle-detected
SDK
aurra@0.5.1 (npm)

See it on a real query

The playground runs against your own memories. Paste an API key, ask a question, and the full proof tree renders inline — retrieval filters, evidence entries, bi-temporal window, source citations, supersession chains.

Open the playground →