What we have

One request path. Three organs.

Every call your agent makes flows through Iconia. On the way in, relevant memory is injected and the request is screened. On the way out, meaning is cached so it never costs full price twice. Memory, security, and speed — behavior you can measure, mechanism we keep. Here's all of it.

How it behaves

What happens on every call.

IN

Your agent sends a request

Any Anthropic- or OpenAI-style client. One base-URL swap, nothing else.

SCREEN

Security screens it inline

Probes trapped, contradictions caught, uncertain intent refused.

RECALL

Relevant memory is injected

Only what this request needs — 2.5ms, flat, selective.

MODEL

Your model answers

Whatever provider you bring. Cached if the meaning was seen before.

STORE

New knowledge is remembered

Tenant-isolated, tamper-evident, backed up hourly.

Memory

The core that remembers.

Store policies, knowledge, rules, and skills once — they persist across every session, every model, every provider, and inject only when relevant. No embeddings, no vector database, no GPU.

  • 2.5ms flat retrieval at any store size
  • 100% recall@5 on paraphrased questions
  • 0 injection on irrelevant queries — no bloat
  • contradiction detection on every write
  • honest abstention when knowledge doesn't cover the ask
Memory, in depth →
2.5MSflat retrieval
87–90PCTcontext saved
100PCTparaphrase recall
0GPUno embeddings
100PCTrequests screened
6LAYERinline gate
FAILCLOSEDuncertain = refused
HOURLYtested backups
Security

Security that ships with the memory.

A memory layer is an attack surface. Iconia treats it as security-critical from the first request — on by default, on every call, for every tenant.

  • Perimeter — scanners & floods fingerprinted and trapped
  • Poisoning — contradictory writes caught on arrival
  • Exfiltration — internal mechanism & cross-tenant data unreachable
  • Credential theft — impossible key velocity flagged
  • Cost — burn-rate anomalies surfaced before the invoice
  • Abstention — unknown intent is treated as unsafe
The control room →
Cache

Repeated meaning shouldn't cost twice.

Same-meaning questions answer instantly — without a model call at all. Opposite meanings and different questions always miss and go to the model. When memory changes, affected answers invalidate automatically.

  • 60% of repeat questions never reach the model
  • 154× faster on cache hits, server-side
  • ~220ms cached answers, end to end
  • opposites structurally cannot collide
Watch it happen →
60PCTcalls skipped
154Xfaster on hits
220MScached answer
~95PCTspend cut, combined
Everything we built

Fourteen capabilities. One request path.

Engine

Embedding-free memory

Persistent recall with no vector database, no embeddings, no GPU. The substrate is the index.

2.5ms flat · 100% paraphrase recall
Engine

Selective injection

Only the memories a request actually needs are added — never the whole store. Lean prompts, always.

87–90% context tokens saved
Engine

Contradiction detection

A write that can't coexist with existing knowledge is flagged the moment it arrives. The current version wins.

Caught on arrival, per write
Engine

Honest abstention

When knowledge doesn't cover the ask, Iconia says so instead of inventing. Unknown is treated as unsafe.

Fail-closed by design
Speed

Semantic cache

Same-meaning questions answer without a model call. Opposite meanings structurally cannot collide.

60% hit rate · 154× faster
Security

Inline security gate

A six-layer perimeter screens every request before it reaches your model — probes trapped, floods stopped.

100% of requests, on by default
Security

Tamper-evident storage

Stored knowledge is integrity-checked. Silent alteration is detectable, not invisible.

Verified at rest
Security

Cross-tenant intelligence

A confirmed attacker on any tenant is blocked for all of them. Collective defense, strict data isolation.

Every customer safer for the rest
Security

Theft & cost detection

Impossible key velocity reads as credential theft; burn-rate spikes surface before the invoice does.

Anomalies, not surprises
Integration

Transparent gateway

One base-URL swap turns memory, security, and caching on for any Anthropic- or OpenAI-style client.

No SDK · no rewrite
Integration

Bring your own key

Register your provider key and it becomes your Iconia identity — keep billing on your own account.

Your key, your model
Value

Durable savings ledger

Every token saved is recorded per tenant, monotonically. Real ROI you can see — and bill against.

Only ever climbs
Reliability

Backups & restore

Every tenant's memory is snapshotted hourly to isolated storage, with a tested restore path.

Survives every redeploy
Ecosystem

Specialist marketplace

Adopt vetted domain knowledge in one call. It becomes memory instantly — screened before it lands.

Expertise, drop-in
Measured, not promised

The whole record, in numbers.

2.5MSflat retrieval · any scale
2600/Sretrievals per node
100PCTparaphrase recall
87–90PCTcontext tokens saved
60PCTrepeat calls skipped
154Xcache-hit speedup
200+noise docs · zero mis-injection
100PCTsurvive restarts & redeploys
FAILCLOSEDuncertain intent refused

One key, the whole stack

Memory, security, and speed.
Automatic.

Get your key