Memory you can verify.

AI agents that remember — and can prove it.

Aurra gives your AI a memory you can actually trust. Every answer comes with a receipt: what it knew, when it knew it, and why it responded the way it did — for teams who can't just take the AI's word for it.

Python & JavaScript SDKs are live. pip install aurraView changelog →

AI agents forget.
And when they remember, you can't see why.

As agents start making real decisions, "the AI remembered it" isn't good enough. You need to know what it knew, when it knew it, and what it based its answer on — especially when something goes wrong.

🔍
No way to verify recall
Most memory layers give you an answer with no trail. When the agent says something wrong, you can't see which memory it pulled, or why. Aurra returns a proof tree with every answer — the exact evidence behind it.
🕒
Memory drifts over time
Facts change. A preference from last year may be stale today. Without a sense of time, agents confidently repeat outdated information. Aurra is bi-temporal — it knows what was true, and when.
🛡️
Nothing to show an auditor
In regulated or high-stakes settings, "trust the model" doesn't pass review. Aurra keeps a tamper-evident record of every memory and every answer — built for the moment someone asks you to prove it.

Built for teams who
need to be sure.

From regulated enterprises to solo builders — wherever an agent's memory has to be trusted, traceable, or simply out of your way.

🏥
Healthcare
Patient-facing AI that holds up to review

A clinical assistant surfaces a fact about a patient — a medication, an allergy, a past visit — and a clinician acts on it. When that fact turns out wrong or outdated, the question is brutal: where did the AI get this, and when? Most memory layers have no answer. They overwrite old facts, lose or fabricate timestamps, and return recall as a black box. “The model said so” does not survive a clinical review.

How Aurra solves it

Aurra returns a proof tree with every answer — the exact memories used, a tamper-evident hash, and what was superseded. Bi-temporal versioning means you can see what was known and when, so an outdated fact shows up as outdated instead of being silently repeated. Every recall carries a full audit trail, built for the moment someone asks you to account for it.

Audit and traceability capabilities — not a substitute for your own compliance certification.
🏢
Enterprise
Agents that make decisions, on the record

An internal agent starts taking real actions — approving, routing, recommending — based on what it remembers about your business and customers. The first time one of those decisions is questioned, you need to reconstruct exactly what the agent knew at that moment and why it chose what it did. Most memory layers keep no decision trail: they cannot tell you which memories drove an answer, whether any were stale, or what changed. Every agent decision becomes an unauditable liability your risk team will not sign off on.

How Aurra solves it

Aurra records the memory behind every answer as a verifiable evidence chain. Bi-temporal queries let you ask “what did the agent believe on this date?” and get a precise answer. Native, database-enforced tenant isolation keeps each business unit or customer cleanly separated. When a decision needs explaining, the record is already there.

🚀
Startups & small teams
A memory layer you ship in an afternoon

You’re building an agent product and your users expect it to remember them across sessions. Building that yourself means a vector store, an extraction pipeline, tenant isolation, and the ops to run it all — weeks of infrastructure that isn’t your actual product. Rolling your own multi-tenant isolation is exactly the kind of thing that breaks quietly and bites the moment you land a B2B customer who asks how their data is separated.

How Aurra solves it

Five lines and you have a working memory layer, in Python or TypeScript. Multi-tenant from day one, with isolation enforced at the database layer, so you can sell to B2B customers without rebuilding anything. Bring your own LLM key so your inference costs stay on your bill — not marked up on ours.

Founders & solo builders
No memory plumbing to build or babysit

You’re one person shipping an agent, and “give it memory” keeps turning into a project of its own. Every hour spent standing up and maintaining a memory system is an hour not spent on the thing that makes your product yours — and once it’s running, it’s one more piece of infrastructure to keep alive at 2am.

How Aurra solves it

Drop Aurra in and your agent remembers across every session — nothing to deploy, nothing to run. The audit trail and proof trees are already there if you ever need to explain a result, debug a bad answer, or show an enterprise customer how recall works. The hard parts are handled; you ship.

Memory with
a paper trail.

Built for developers shipping production agents. Verifiable, bi-temporal, multi-tenant. Bring your own LLM — your costs stay on your bill.

01
Proof tree on every answer
Ask a question and get the answer plus a structured record of how it was reached: a tamper-evident hash, the exact memories used, and what was superseded. When recall is wrong, you can see why.
Live in API today
02
Bi-temporal versioning
Track when a memory was captured and when it was true. Query the state of memory at any point in the past. When facts change, Aurra preserves the timeline instead of overwriting it.
Live in API today
03
Source citations + audit trail
Every memory traces back to where it came from, with a full audit endpoint — extraction model, prompt version, and event history. Built for the moment someone asks you to prove it.
Live in API today
04
Python + JavaScript SDKs
Install in five lines, in either language. A clean, typed client for query, write, audit, timeline, and proof trees. Migrate or start fresh in minutes.
pip install aurranpm install aurra
05
Multi-tenant native
Every write takes a tenant_id for end-user scoping, enforced at the database layer. Build B2B agents without rolling your own isolation.
Live in API today
06
Bring your own LLM
Pass your own provider key per request — Anthropic, OpenAI, your choice. Your API costs stay on your bill. Or use the Aurra LLM add-on to extract memories for you at $0.005 per write.
Anthropic ✓OpenAI ✓Aurra LLM ✓

Five lines to
memory with receipts.

Add trustworthy memory to your agent in three steps — Python or TypeScript, no infrastructure to provision.

1
Install the SDK
pip install aurra (or npm install aurra). Get an API key from your dashboard. Nothing to deploy or run yourself.
2
Write memories from your agent
Pass conversation turns or facts to add(). Aurra extracts the meaningful memories, preserves source timestamps, and scopes them by tenant.
3
Query — with proof
query() returns an answer plus the memories behind it. Ask for a proof tree to get the full evidence chain: what was retrieved, what was superseded, and a tamper-evident hash.

Numbers we measure.
Methodology we publish.

Every benchmark we cite is reproducible. Raw hypothesis files, judge prompts, and methodology are open at github.com/aurra-memory/benchmarks.

01
LongMemEval-S: 80.2% mean
432 of 500 questions, K=30 retrieval, two-judge validation. GPT-4o judge: 81.3%. Claude judge: 79.2%. Cross-judge agreement within 2pp confirms the result is robust.
Our weakest category is single-session-assistant at 33.9%. We disclose it — retrieval reranking is in progress.
02
LoCoMo: source timestamps preserved
On 2,685 memories from 10 LoCoMo conversations, Aurra stamped zero with fabricated dates — every memory keeps the timestamp of the conversation it came from. Deterministic count, not LLM-judged. Methodology and raw data published.

Pre-built memory ingestion adapters for common sources. Use them as examples — or write your own with the SDK.

💬SlackLive
📝NotionLive
📧GmailComing soon
📹ZoomComing soon
🗂️LinearComing soon
📋JiraComing soon

Pick a plan.
Cancel anytime.

Free for prototypes. Pay when you ship. Enterprise when compliance asks.

Hobby
$0 forever
Prove it works in your stack.
  • 500 writes / month
  • Both SDKs (Python + JS)
  • Proof trees + source citations
  • Audit endpoint
  • BYO-LLM
  • 1 tenant
Start free →
Starter
$29 / month
When your agent has real users.
  • 50,000 writes / month
  • Both SDKs
  • Multi-tenant native
  • Proof trees + audit
  • BYO-LLM
  • Email support
  • 14-day free trial
Start trial →
Enterprise
Custom
When compliance asks questions.
  • Unlimited writes
  • SSO / SAML
  • Custom connectors
  • Dedicated Slack
  • Audit log export + DPA
  • Self-hosted deployment (roadmap)
Contact sales →

All plans billed monthly in USD. See full plan details →

Give your agent a memory
you can trust.

Five lines to a memory layer with a proof tree on every answer. Free to start — no infrastructure to run.

Start free →

pip install aurra · npm install aurra