ICONIA

Give your AI a self.

Save time and money. Your agent remembers your work — across every session, every model.

Start free

Works with Claude, GPT, Gemini, Llama, and local models.

See your savings
tokens saved across every agent
Works with every model you already run
OpenAIAnthropicGeminiClaudeOpenRouterOllamaCursorHermesWindsurfLocal LLMsSelf-hosted
The problem

Your AI isn't actually learning.

Users repeat themselves

Every conversation starts from zero — the context you already paid for is gone.

Agents forget context

Preferences, projects, prior decisions — dropped the moment the session ends.

You pay twice

The same tokens re-sent on every call. The bill grows with your success.

Every prompt is exposed

A raw prompt to an LLM is an open door for injection, jailbreaks, and abuse.

The more successful your product becomes, the worse this gets.

Get going

Add memory in three steps.

Point the tool you already use at Iconia and memory, security, and caching switch on for every call. No SDK, no rewrite — you change one line.

01

Get your key

Sign up and copy your ick_ key — shown once.

02

Paste it into your tool

Set one base URL in Cursor, Claude, or any OpenAI-style app — pick yours below.

03

It just remembers

Teach it something now; a new session recalls it — automatically. Watch it live on your dashboard.

# Cursor → Settings → Models → OpenAI API — toggle a custom base URL
Base URL:  https://www.iconia.tech/v1
API Key:   ick_your_key

Paste those two lines and every message your agent sends stores and recalls memory — nothing else to change.

# Claude Code — set two environment variables
ANTHROPIC_BASE_URL=https://www.iconia.tech/v1
ANTHROPIC_API_KEY=ick_your_key

Run claude exactly as before — it remembers across every session, no other change.

# Hermes — config.yaml (or set OPENAI_BASE_URL / OPENAI_API_KEY)
model:
  provider: custom
  base_url: "https://www.iconia.tech/v1"
  api_key:  "ick_your_key"

Hermes is model-agnostic and self-hosted — route its model through Iconia and your always-on agent gets governed, portable memory + security on every call. Same "bring your model, keep your memory."

from openai import OpenAI

client = OpenAI(
  base_url="https://www.iconia.tech/v1",
  api_key="ick_your_key",
)

Drop-in for any OpenAI-compatible app or script — just these two fields.

Get your key
Measured, not promised

The numbers are the pitch.

1MSflat retrieval · any store size
87–90PCTcontext tokens saved per request
60PCTrepeat questions never reach the model
220MScached answers · end to end
154Xfaster on cache hits · server-side
One request path

What happens on every call.

Request→ Security screens→ Memory injects→ Cache checks→ Model answers→ ROI measured→ Knowledge stored
One request path

Everything flows through, and comes back smarter.

Memory

It remembers

Your policies, knowledge, and rules — injected into every call that needs them. Your agent stops asking you to repeat yourself.

How it behaves
Security

It guards

Every request screened inline — probes trapped, contradictions caught, uncertain intent refused. Security that ships with the memory.

The control room
Cache

It compounds

Repeated meaning should not cost full price twice. Same-meaning questions answer instantly — without a model call at all.

Watch it happen
The control room

See what it's doing — and what it saved.

Tokens saved—
Requests served—
Model calls avoided—
Retrieval, flat1ms
Cache hit rate60%
Security screening100%
BackupsHourly
Tenant healthOperational
Open the dashboard
Who it's for

Built for teams shipping agents.

AI agent developers
IDE & coding-agent builders
SaaS teams running LLMs in production
Support-automation teams
Enterprises deploying governed AI
Security-conscious AI infra teams
Build it yourself, or ship today

Weeks of infrastructure, or one endpoint.

Build it yourself Iconia
Weeks or monthsMinutes
Vector DB + Redis + pipelines to stitchOne endpoint
Ongoing maintenanceFully managed
Security left to youBuilt-in Sentinel
High token wasteOptimized + cached

Memory shouldn't be a feature you maintain. It should be infrastructure you call.

The questions you'd ask

Answered up front.

Won't it bloat my prompts?
The opposite. Ask about something irrelevant and zero is injected — no wasted tokens, no noise. Only what the request actually needs arrives, capped at five items. Bloat is what you're leaving behind.
How much does it actually save?
87–90% of context tokens per request, and up to 60% of repeat questions never reach the model at all. On repeat-shaped traffic that's most of your bill — gone.
Will it slow my agent down?
Retrieval runs in under 10ms and stays flat no matter how much you store — measured with the right memory still ranked first at 1,000 stored items. Cached answers return in ~220ms, faster than the model could ever reply. It makes agents quicker, not slower.
Do I have to rewrite anything?
No SDK, no migration. Swap one base URL and memory, security, and caching switch on for every call. Works with any Anthropic- or OpenAI-style client you already run.
What if two rules conflict?
Caught the moment the second one arrives. The current version is served; the contradiction is flagged for review. Your agent never confidently tells two people two different things.
Am I locked in?
Never. Your memory is tenant-isolated and portable, your model is whatever you bring, and you can leave any time. The value accrues to you — not to a format you can't export.
Two lines

If you can set a base URL, you're done.

Point your existing client at Iconia. Memory, security, and caching switch on for every call — no SDK, no rewrite, no new habits.

# your client, unchanged — plus memory
client = OpenAI(
  base_url="https://www.iconia.tech/v1",
  api_key="ick_your_key"
)

Knowledge that remembers

Bring your model.
Keep your memory.

Get your key