Immortal Virtual Agent Sessions

Iter Verum Ad Singularitatem — The True Journey to Singularity

IVAS

AI that doesn't forget — and doesn't make things up.

Your AI agent remembers everything. Forever. Across crashes, restarts, and updates. True immortal memory — not just "persistent storage."

Start Free Trial Try the Einherjar Protocol

3 days free, full access. No credit card required.

93.7%

Adversarial Recall

Click for details

81.2%

LongMemEval Overall

Click for details

Einherjar Protocol

Click for details

Dependencies

Click for details

<1ms

Recall Latency

Click for details

LoCoMo Benchmark — 93.7% Adversarial

93.7%

What is LoCoMo?

The Long-term Conversational Memory (LoCoMo) benchmark evaluates whether an AI system can remember facts, preferences, and events from long conversation histories — even after hundreds of exchanges.

What it tests

Factual recall — "What did the user say about their job 3 months ago?"
Preference tracking — "Does the user prefer dark mode or light mode?"
Temporal reasoning — "When did they first mention this topic?"
Multi-session continuity — Information from session 1 must survive to session 50+

How IVAS achieves this

Every memory is categorized, timestamped, and instantly searchable. IVAS uses a proprietary persistence engine with crash-proof storage. Nothing is ever lost — not to context window limits, not to crashes, not to restarts.

Official LoCoMo Results (1,986 questions)

Category	IVAS Score	Questions
Adversarial (won't hallucinate)	93.7%	446
Single-hop recall	53.7%	841
Temporal reasoning	39.9%	321
Multi-hop reasoning	33.7%	282
Overall	55.4%	1,986

Evaluated using official LoCoMo F1 scoring with Claude Haiku for extraction.

Why some scores are lower

IVAS is truth infrastructure — we store and retrieve, we don't reason. High adversarial score = we don't hallucinate. Lower reasoning scores = that's your LLM's job. Same IVAS, different models, different reasoning scores. The memory layer is constant.

LongMemEval — 81.2% Overall (Opus 4.6)

81.2%

What is LongMemEval?

LongMemEval is a benchmark specifically designed to test long-term memory in conversational AI systems. It measures whether information stored weeks or months ago can be accurately retrieved.

What it tests

Information extraction — Can the system extract and store key facts from conversations?
Knowledge update — Can it update previously stored information when corrected?
Knowledge application — Can it use stored memories to answer future questions?
Cross-session persistence — Does information survive session boundaries?

How IVAS achieves this

IVAS stores every memory with full metadata (category, subject, confidence score, timestamp). Our proprietary search engine finds relevant memories in under 1 millisecond regardless of vault size. Corrections are tracked separately — when information updates, the old version is preserved and the new version takes priority.

Official LongMemEval Results — BYOM Model Comparison

Same IVAS memory layer. Different models. Proof that memory is the bottleneck, not the LLM.

Model	Overall	KU	User	Asst	Pref	Multi	Temp
Opus 4.6	81.2%	100%	97.1%	78.6%	76.7%	78.2%	66.9%
Sonnet 4.6	80.2%	96.2%	98.6%	78.6%	70.0%	75.9%	68.4%
Haiku 4.5	71.8%	96.2%	95.7%	76.8%	50.0%	60.9%	58.7%
GPT-4o	71.4%	96.2%	88.6%	69.6%	50.0%	60.9%	63.9%
GPT-4.1-mini	70.6%	92.3%	98.6%	76.8%	50.0%	58.7%	57.1%
GPT-4o-mini	66.2%	88.5%	97.1%	75.0%	40.0%	51.1%	54.1%

500 questions each. GPT-4o-mini as judge. KU = Knowledge Update. User/Asst/Pref = Single-session categories. Multi = Multi-session. Temp = Temporal reasoning.

Same memory. Different models. Different scores.

All 6 models use identical IVAS memory. The only variable is the LLM. Knowledge Update is 96-100% across all models because that's pure memory. Multi-session and Temporal vary because that's reasoning. IVAS provides the facts. Your model interprets them.

Key advantage

Most systems use vector embeddings for memory retrieval, which degrades over time (embedding drift). IVAS uses exact matching — the same query returns the same result whether you run it today or in 5 years. No drift. No degradation. Ever.

One-Line Install

What this means

pip install ivas — that's it. No MongoDB. No Redis. No Docker. No cloud setup. No configuration files. One command and you're connected to your private memory vault.

Why this matters

System	Setup required	Time to first memory
IVAS	pip install ivas	30 seconds
Mem0	Qdrant + embeddings model	10+ minutes
Zep	Postgres + Redis + Docker	30+ minutes
LangChain Memory	Vector DB + embeddings	15+ minutes

Install IVAS. Get your API key. Three lines of code. Your agent remembers.

Sub-1ms Recall Latency

<1ms

What this means

When your agent asks "what does the user prefer?", the answer comes back in under 1 millisecond. No network round-trips. No embedding calculations. No API calls. Instant.

Why it's so fast

Optimized infrastructure — Your vault runs on dedicated, performance-tuned servers.
Pre-built indexes — Queries hit optimized indexes, not raw data scans.
No embeddings — Competitor systems add 10-50ms per query for vector similarity. IVAS uses exact matching — faster by orders of magnitude.

Comparison

System	Recall latency	Bottleneck
IVAS	<1ms	None (optimized)
Mem0 Cloud	50-200ms	Network + embedding
Pinecone	20-100ms	Network + ANN search
Zep Cloud	30-150ms	Network + graph query

For voice-first AI agents, sub-1ms memory recall is the difference between a natural conversation and an awkward pause.

Other AI memory is not immortal

Every AI agent today is born with amnesia. Context windows fill up. Sessions crash. Tokens expire. And the agent forgets everything. You start over. Every. Single. Time.

"Persistent Memory"

What competitors actually offer:

Memories stored in the cloud (their servers, not yours)
Lost during crashes or server outages
Dependent on embedding models that drift over time
No verification that memories survived
Requires Docker, Redis, MongoDB, cloud services

Immortal Memory (IVAS)

What we actually deliver:

Your data in your private vault — isolated infrastructure
Survives crashes, restarts, updates — proven, not promised
Exact matching — no drift, no degradation, ever
Einherjar Protocol verifies nothing was lost
One-line install — pip install ivas and you're connected

The Einherjar Protocol

Named after the Norse warriors who died in battle and were reborn in Valhalla with every memory intact. Your AI agent doesn't just survive death — it comes back whole.

How it works — in plain English

Imagine your AI agent is working on a task. It knows your name, your preferences, what you discussed last week, and what it was about to do next. Then it crashes.

With other systems, some of that might be saved. Maybe. You'd never know what was lost.

With IVAS, the Einherjar Protocol does three things:

Snapshot — Before anything happens, IVAS records everything the agent knows
Resurrect — After a crash, IVAS restores the agent from its dual-write storage
Audit — Einherjar compares the restored state against the snapshot and generates a verification report

The result? A coverage percentage. 100% means every single piece of knowledge survived. That's S-rank.

S = 100% knowledge verified. The highest possible rating. Nothing was lost.

S = 100% (Perfect) A = 90-99% B = 75-89% C = 50-74% F = Below 50%

Try Einherjar Yourself

Run the protocol in your browser — no install required

Truth Infrastructure

IVAS doesn't think. IVAS doesn't interpret. IVAS doesn't decide what matters. That's your AI's job.

What IVAS does

Stores exactly what you give it — unchanged, unfiltered
Returns exactly what you stored — no reinterpretation
Verifies nothing was lost — provable, auditable
Keeps corrections separate from originals — full history preserved

What your AI does

Reasons across memories — connects the dots
Interprets context — understands meaning
Makes decisions — applies your rules and logic
Learns from corrections — grows smarter over time

Your AI is the brain. IVAS is the memory.

We don't touch your data. We don't train on it. We don't filter it. If you store garbage, we save garbage faithfully. If you store gold, we save gold faithfully. That's the promise.

This separation is intentional — and it's why our benchmark scores tell a specific story.

Reading our benchmarks

Some of our scores are high. Some are intentionally lower. Here's what each one measures and why.

Adversarial Recall 93.7%

What it means: When IVAS doesn't have the answer, it says so instead of making something up. This is memory integrity — we never hallucinate facts.

This is us. High score = we don't lie.

Knowledge Update 100%

What it means: When you correct a fact, the correction sticks. Every single time. Old info preserved, new info takes priority.

This is us. Perfect score = corrections always work.

Single-hop Recall 53.7%

What it means: Finding one specific fact you stored. Part memory retrieval (us), part LLM extraction (your model). Score varies by model.

Shared responsibility. We find it, your AI reads it.

Multi-hop Reasoning 33.7%

What it means: Connecting two separate facts to answer a question. This is pure reasoning — your LLM's job, not ours. We deliver the facts. Your model connects them.

This is your AI. Varies by model. Not our scope.

Temporal Reasoning 39.9%

What it means: Understanding when things happened and in what order. This requires reasoning about time — your LLM's domain. We timestamp everything. Your model interprets the timeline.

This is your AI. We provide timestamps, your model reasons.

BYOM 6 models

What it means: Bring Your Own Model. Same IVAS memory, any AI. Plug in Claude, GPT, Gemma, LLaMA — the reasoning scores change, the memory scores don't. Proof that memory and intelligence are separate layers.

Your choice. Your model. Our memory.

How IVAS works

Get immortal memory for your AI agent in 3 steps. No coding required.

Sign Up

Create your account on ivas.dev and get your personal API key. Your key unlocks your private memory vault — encrypted and isolated on its own server.

Connect

Point your AI agent to the IVAS API with your key. Works with any agent framework — Claude, GPT, LLaMA, or custom. One line of configuration.

Never Forget

Your agent now remembers everything, forever. Crash it, restart it, update it — every memory survives. Verified by the Einherjar Protocol.

Python SDK. One line to install. Three lines to immortal memory.

Install via pip:

            pip install ivas
            Copied!
        

from ivas import Client

# Connect to your private vault
client = Client(api_key="your-key")

# Store memories — survives any crash
client.remember("user", "preference", "Boss prefers dark mode")
client.recall("dark mode")  # Returns instantly (<1ms)

# Session continuity — picks up mid-thought
client.save_session(
    active_task="Processing invoices",
    current_thought="Checking line 47 for tax errors",
)
# After crash: client.restore_session() → right where you left off

# Verify nothing was lost
report = client.audit()  # 100% coverage = S-rank
        

CLI also available:

$ export IVAS_API_KEY="your-key"
$ ivas remember user preference "Prefers dark mode"
Stored memory #1
$ ivas recall "dark mode"
[1] user/preference: Prefers dark mode (conf=0.8)
$ ivas stats
Vault: my-agent | Memories: 147 | Corrections: 3 | Sessions: 12
        

Your data. Your vault. Your keys.

Every IVAS customer gets their own isolated, private memory vault. Your data never touches another customer's server.

🔒

Private Vault

Each customer gets an isolated database on dedicated infrastructure. Your memories are physically separated from every other customer.

🔑

Your API Keys

Access your vault through encrypted API keys that only you control. Keys are hashed server-side — even we can't see them. Revoke or rotate anytime.

🛡

Zero Shared Infrastructure

No shared databases. No multi-tenant vector stores. No "your data helps improve our model." Your agent's memories are yours alone.

IVAS vs. the rest

Feature	IVAS	Mem0	Zep	Others
True crash recovery	Verified (Einherjar)	Not verified	Not verified	No
Knowledge verification	Built-in audit	None	None	None
Data sovereignty	Private VPS vault	Shared cloud	Shared cloud	Varies
Setup complexity	pip install	Qdrant + model	Postgres + Redis	Various
Recall latency	<1ms	50-200ms	30-150ms	20-200ms
Managed hosting	Fully managed	Self-manage	Self-manage	Varies
Patent-pending	64/019,813	No	No	No

Early Access Pricing

Lock in Early Access pricing today. Rates increase after launch.

Limited to 200 Early Access Vaults

Once 200 vaults are claimed, pricing shifts to full price. Early Access members lock in their discounted rate permanently.

Trial

Free / 3 days

Full access for 72 hours. See for yourself.

Zero Memory Loss Guarantee: If your agent loses a single memory during the trial, get 6 months free.

1 agent vault
1,000 memories
500 API calls
Einherjar verification
Developer SDK included
Auto-expires after 72 hours

Operator

$99$49/mo

For individuals and small teams building AI agents.

3 agent vaults
50,000 memories
10,000 API calls/mo
Private VPS partition
<1ms recall latency
Crash-proof sessions
Einherjar verification
Email support
Early Access price locked in

Studio

$179$99/mo

For teams running multiple agents at scale.

10 agent vaults
500,000 memories
100,000 API calls/mo
Dedicated VPS instance
Priority support
Custom retention policies
Webhook integrations
Team management
Early Access price locked in

Enterprise

Custom

For organizations with compliance and scale requirements.

Unlimited agent vaults
Unlimited memories
Unlimited API calls
Fully sovereign VPS
SLA guarantee
HIPAA-ready configuration
Dedicated onboarding
On-premises deployment option
Annual licensing available

Start Free Trial

3 days. Full access. No credit card.

or sign up with email

Make your AI agent immortal

3-day free trial. Full access. No credit card required.
Your agent will never forget again.

Start Free Trial Try Einherjar Protocol

Patent Pending: 64/019,813 (Provisional)

AI that doesn't forget — and doesn't make things up.

LoCoMo Benchmark — 93.7% Adversarial

What is LoCoMo?

What it tests

How IVAS achieves this

Official LoCoMo Results (1,986 questions)

LongMemEval — 81.2% Overall (Opus 4.6)

What is LongMemEval?

What it tests

How IVAS achieves this

Official LongMemEval Results — BYOM Model Comparison

Key advantage

Einherjar Protocol — S Rating

What does "S" mean?

What is the Einherjar Protocol?

How it works

Why this matters

One-Line Install

What this means

Why this matters

Sub-1ms Recall Latency

What this means

Why it's so fast

Comparison

Other AI memory is not immortal

"Persistent Memory"

Immortal Memory (IVAS)

The Einherjar Protocol

How it works — in plain English

Truth Infrastructure

What IVAS does

What your AI does

Reading our benchmarks

How IVAS works

Sign Up

Connect

Never Forget

Your data. Your vault. Your keys.

Private Vault

Your API Keys

Zero Shared Infrastructure

IVAS vs. the rest

Early Access Pricing

Trial

Operator

Studio

Enterprise

Start Free Trial

Make your AI agent immortal