✨ Permanent AI Memory · 60-Second Setup · Zero Hallucinations →
by Calera Labs

Give Your AI An Unforgettable Memory.
Zero Hallucinations. 50% Cheaper.

Stop prompt stuffing. Stop paying for forgotten context. Calera gives your agents and LLMs permanent, instant memory that never degrades, never hallucinates, and cuts your token bills in half. Plug into Cursor, Claude, or Python in 60 seconds.

Start Free Pilot (1M Nodes Free) Try Interactive Demos
No credit card required • 1-click Cursor & Claude setup • Drop-in OpenAI SDK
🤗 New on Hugging Face • Bedrock Models Built on Bedrock. Verified by ICX.

How ICX Becomes the Hero for Open-Weights Models: 77.0% Alone → 99.23% With ICX

Running our open-weights PatentSec-12B (GGUF) model locally in Ollama or llama.cpp? Alone, a 12B model scores 77.0% (218/283) on PatentBench v0.4.0 Mini (96.5% on self-contained prompts, 0/19 on unseen USPTO dockets). Pair it with the ICX Epistemic Memory Substrate and ICX rescues every blind spot — lifting Mini accuracy to 92.6% (+44 cases) and delivering 99.23% (7,042 / 7,097 verified passes) across the full enterprise corpus with 0.00% confabulation (∂² = 0).

Hugging Face Weights Alone CaleraLabs/patentsec-12b-gguf
77.0% Layer 1 Mini (218/283)
96.5% self-contained • 0/19 unseen file wrappers • No live docket memory
⚡ ICX as the Hero (Weights + Substrate) icx.api.caleralabs.com
99.23% Full Corpus (7,042/7,097)
92.6% Mini (+15.6% lift) • 18/19 examiner recall • 0.00% hallucination (∂² = 0)
Sign Up Free — Make ICX Your Model's Hero (1M Nodes) 🤗 Hugging Face Weights (GGUF) ↗ ⚖️ Test 77% vs 99% in HF Arena ↗ 🏛️ Open PatentSec Portal →
Standard AI (Without Memory) Amnesic & Expensive
"What were our Q2 billing rules discussed 3 weeks ago?"
"I apologize, but I don't retain memory from earlier conversations. Could you paste the memo again?"
45,000 Prompt Tokens
4.8 s Prefill Latency
Context Rot Recall Quality
With Calera ICX Instant & Grounded
"What were our Q2 billing rules discussed 3 weeks ago?"
"Per your team memo on May 14th, Q2 billing enforces net-30 payment terms with a 2% prompt-pay discount."
120 Prompt Tokens
0.2 s Instant Recall
100% Match Proven Grounding
50%+
Prompt Token Savings
51.98% mean token cut on LongBench v2 (503 tasks). Stop re-sending massive prompt histories.
< 2 ms
Instant Memory Recall
Sub-millisecond topological retrieval on ordinary CPU. Instant answers with zero GPU prefill lag.
96.3%
Needle Retrieval Precision
466/484 exact retrieval on RULER MRCR multi-round test. Retains precise facts across unbounded context.
0.00%
Hallucination Guarantee
Mathematical boundary firewall (∂² = 0). Guaranteed 100% verified facts or honest refusal.
Why Builders Choose Calera

Permanent Memory. Zero Hallucinations. 50% Lower Bills.

Stop stuffing massive documents and chat histories into expensive LLM prompts. Store your knowledge once, recall it instantly, and give your AI an unforgettable memory.

🧠

It Never Forgets

Continuous memory across files, chats, and workflows

Your AI keeps full context across days and weeks without starting from scratch. No lost-in-the-middle context rot, no forgotten instructions, and no amnesia when conversations grow long.

  • Persistent state across multi-turn sessions
  • Auto-syncs codebases, memos, and files
  • Works with Cursor, Claude, and OpenAI SDK
Mathematically Proven
🛡️

The Hallucination Firewall

100% verified facts or honest refusal

Built on discrete topological boundary math (∂² = 0), Calera guarantees that if a fact is not in your verified knowledge vault, your model won't fabricate a convincing lie. Perfect for mission-critical apps.

  • 0.00% statistical confabulation
  • Cryptographic SHA-256 provenance chains
  • Audit-ready citations down to the exact sentence
⚡

Cut Token Bills by 50%+

Instant sub-2ms recall on ordinary CPU

Instead of re-sending 50,000 to 100,000 prompt tokens every single turn, Calera pinpoints the exact 500 tokens needed and streams them instantly to your LLM. Massive cost reduction and 4x faster replies.

  • 51.98% mean token cut on LongBench v2
  • Sub-millisecond retrieval latency
  • Zero expensive GPU prefill overhead
Empirical Head-to-Head Evaluation

Head-to-Head: LLM Alone vs. LLM with Calera ICX

Paired evaluation used Google Gemini 3.5 Flash-Lite (gemini-3.5-flash-lite) as the upstream BYOK test model. ICX works with any supported LLM — Gemini was the model we ran for this comparison. Same tasks, same weights — ICX separates permanent lattice storage from the generative viewport so the model receives only the facts it needs. Token cuts below are versus uncached full-context Gemini prefill. They are not a dollar comparison to a prompt-cached Cursor, Claude Code, or Antigravity harness.

Upstream test model Google Gemini 3.5 Flash-Lite · API ID gemini-3.5-flash-lite · BYOK via Google AI API
LLM alone Gemini 3.5 Flash-Lite · full context prefill
LLM + Calera ICX Gemini 3.5 Flash-Lite · scoped viewport

Open-weights statutory reasoning (Hugging Face PatentSec-12B)

PatentBench v0.4.0 · 283-case Layer 1 Mini ablation & 7,097-case Full Enterprise Corpus

PatentSec-12B (HF GGUF / bf16) alone
77.0% (218/283 Mini)
96.5% self-contained • 0/19 unseen examiner dockets
PatentSec-12B + Calera ICX
99.23% (7,042/7,097 Full · 92.6% Mini)
+44 Mini cases rescued • 18/19 examiners • 0.00% confabulation (∂² = 0)
+15.6% Mini lift • 99.23% on 7,097 cases • Sign up free to attach ICX to your Hugging Face model →

Long-context reasoning

LongBench v2 paired control · n=10 · identical item IDs

Gemini 3.5 Flash-Lite alone
6/10 (60.0%)
2,253,544 prompt tokens
Gemini 3.5 Flash-Lite + ICX
6/10 (60.0%)
162,923 prompt tokens
Same quality · 92.77% fewer tokens vs uncached prefill · n=10 only

Needle retrieval

RULER MRCR · 484 rows · multi-round ordinal disambiguation

Gemini 3.5 Flash-Lite alone
No 484-row dump
Gemini-alone full MRCR file is not in the public archive
Gemini 3.5 Flash-Lite + ICX
96.28% (466/484 exact)
Wilson 95% CI: 94.2%–97.6% · 18 ordinal sibling misses
466/484 exact · miss ledger closes the rate

Official LongBench v2 (503)

ICX run 073110 · Gemini full-context 20260819_104413 · same 503 IDs

Gemini 3.5 Flash-Lite alone
38.57% (194/503 raw)
188 API failures · valid 194/315 (61.59%)
Gemini 3.5 Flash-Lite + ICX
54.87% (276/503 exact)
0 errors · Wilson 50.5–59.2 · 51.98% mean token cut
+16.30 pp raw · ICX finishes all 503 · Gemini overflows 188

Code repository comprehension

852k context · deep AST & cross-file · item 66fa208b

Gemini 3.5 Flash-Lite alone
6,227 ms (hit)
852,861 full prompt tokens
Gemini 3.5 Flash-Lite + ICX
1,375 ms (hit)
31,783 tokens · 4.53× faster
One matched row · 4.53× · 88.81% vs uncached prefill
View full scorecard table
Benchmark & workload Gemini 3.5 Flash-Lite alone Gemini 3.5 Flash-Lite + ICX Advantage
PatentBench v0.4.0 (HF PatentSec-12B)
283 Mini / 7,097 Full Corpus		 77.0% (218/283 Mini) • 0/19 unseen file wrappers 92.6% Mini (262/283) • 99.23% Full (7,042/7,097) +15.6% Mini lift (+44 cases) • 0.00% confabulation
LongBench v2 paired
n=10 matched item IDs		 6/10 · 2,253,544 tokens 6/10 · 162,923 tokens 92.77% vs uncached prefill · n=10 only
RULER MRCR
484 rows · eval 9fc0e405		 No 484-row Gemini dump in archive 96.28% (466/484) 18 ordinal sibling misses
LongBench v2 official
503 items		 38.57% (194/503 raw) · 188 API failures 54.87% (276/503) +16.30 pp raw · 0 ICX errors
SWE-bench Verified
500 · predicted pass		 No official test-resolved column 96.40% (482/500) Not official test-resolved
Code repository
852k context		 6,227 ms · 852,861 tokens 1,375 ms · 31,783 tokens 4.53× faster · 88.81% token cut
Legal & Compliance
Contract Redlines & MSAs
Track indemnity, liability, and amendment clauses across dozens of contract revisions without ordinal confusion between versions.
Software Engineering
Massive Code Repositories
Traverse AST dependency lattices and cross-file references across monorepos that exceed standard context windows.
Customer Support & CRM
Multi-Turn Ticket Vaults
Keep years of customer dialogue in persistent memory instead of re-uploading full ticket history on every agent turn.
Deep Document Search
Needle-in-a-Haystack Recall
Retrieve exact facts from deep document haystacks without lost-in-the-middle attention degradation.

Subscription Tiers

Select a developer preview tier. All subscriptions include single-account management and multi-domain billing on Calera Labs.

How ICX Persistent Memory Capacity Works

Your subscription tier grants persistent memory capacity (Lattice Nodes) in the Volumetric Lattice Network. Documents stay crystallized in your tenant lattice. On each prompt, ICX projects a numbered viewport of Grounded Facts into your LLM. Token savings depend on the workload (LongBench v2 mean noise cut on 503 rows is 51.98%; synthetic haystack gates are higher).

ICX plans ($0 / $29 / $149 / $999) buy memory capacity. FinanceSec plans are a separate SEC-filing product. One Calera account can subscribe to both — they do not share a token pool.

Monthly Billing Annual Billing Save 20%
Developer Preview

Developer Pilot

For individual developers testing persistent memory workflows.

$0/mo
  • 1,000,000 persistent Lattice Nodes (~2,000 source pages / 1–2 git repos)
  • 1 Active API Key limit
  • 1 Isolated Memory Workspace
  • Projects ~500–1,000 Grounded Facts per turn
  • 30 requests/minute burst
  • OpenAI SDK & Hosted MCP Drop-In
Start Free Pilot
Most Affordable Paid

Builder

For solo developers and agents who want persistent memory without a team invoice.

$29/mo
  • 50,000,000 persistent Lattice Nodes (~100,000 pages / 5–10 repos)
  • 3 Active API Keys
  • Isolated memory workspaces
  • 50K calls / day • <0.5ms lattice diffusion
  • 60 requests/minute burst
  • Standard email support
Start Builder
Developer Preview

Team

For engineering teams building memory-augmented AI applications and swarms.

$149/mo
  • 500,000,000 persistent Lattice Nodes (~1M pages / monorepos)
  • 10 Active API Keys
  • Shared team memory spaces & living auto-sync
  • Multi-tenant namespace isolation
  • 150 requests/minute burst
  • Priority Email + Slack support
Upgrade to Team
Production

Scale

For production AI workloads, heavy context pipelines, and autonomous agent swarms.

$999/mo
  • 5B Lattice Nodes / mo engine allotment (~10M+ source pages)
  • Unlimited applications & workspaces
  • Multi-tenant enterprise isolation
  • ~2M API calls / day • 99.95% uptime SLA
  • Dedicated routing & priority queue
  • Dedicated Slack channel & engineering support
Upgrade to Scale
Dedicated Cluster

Enterprise

Self-hosted + unlimited + white-glove VPC/on-prem clusters. Contact sales.

Custom
  • Unlimited Lattice Nodes & calls
  • Self-hosted / VPC on-prem license
  • Hardware attestation & SOC-2 compliance
  • Dedicated forward-deployed engineer
  • Custom SLA & SAML / Okta SSO
  • Contact sales for customized deployment
Contact Sales
TCO Calculator

Modeled token-volume sketch

A formula, not an invoice. It applies a published list price and a published cache-read multiplier. ICX has not measured a live agent-loop bill.

Real-World Scale Presets:
Real-World Benchmark Representation

SaaS Engineering Team (Team Tier — 5,000,000 Lattice Nodes): Equivalent to 100s of product specs, 50 enterprise MSAs, full API documentation, and 6 months of customer support logs (~10,000 source pages / ~1,250,000 Grounded Facts) crystallized in persistent memory.

Assumptions (all modeled): input $2.50 / MTok; Anthropic cache-read 0.10× if the entire archive were a stable prefix; ICX viewport 1,000 tokens/turn + the matching subscription. A prefix above ~1M tokens does not fit in one window, so the cache-read column is algebra, not a feasible agent turn.

Modeled uncached full replay
$187,500 / mo
Archive × $2.50/MTok × monthly turns. Not how a cached 2026 agent is billed.
Modeled 0.10× cache-read
$18,750 / mo
Published Anthropic hit multiplier on the same token volume. Assumes 100% hits and a prefix that large — usually false.
Modeled ICX viewport + subscription
$67 / mo
1,000 tokens/turn at $2.50/MTok plus the matching plan. Viewport size is an assumption, not the 503-row mean.

No savings percentage is shown. This sketch is not a measured invoice and is not a win against a warm cache.

Technical Whitepaper

The Epistemic Architecture & Benchmark Sweep

Read Infinite Memory Is Not Infinite Attention (CALERA-PAPER-ICX-2026-08-28). Discover the A₄ Simplicial Transitive Lattice Walker (943µs 5-hop), In-Process WASM Kernel (37µs), Multi-Agent Swarms ($V_t$), Universal Auto-Sync, NIST FIPS 203/204 Post-Quantum Security, 15 Certified MCP Tools, and 100% benchmark sweep.

Read the Research Whitepaper
A₄ Transitive Walker: 943µs 5-Hop Causal Walk
In-Process WASM Kernel: 37µs Execution
Multi-Agent Swarm Variable State (V_t) & CAS
Universal Auto-Sync & O(1) Instant Memory Purge
Post-Quantum NIST FIPS 203 (ML-KEM) / FIPS 204 (ML-DSA)
We asked Antigravity

What changes in the agent loop when recall is indexed

Calera asked Google Antigravity, with the ICX MCP attached to this repository, to contrast truncated grep with lattice lookup. The result is an operator transcript — not a measurement. Empirical head-to-head numbers stay on the paper above.

Read the session
Instrument: Google Antigravity + ICX MCP
Genre: operator transcript
Not a live telemetry probe
Companion to the ICX docs
The Substrate Separation Principle

The Desk-and-Warehouse Principle: Decoupling the AI Memory Substrate

Prompt caching discounts re-reading a stable prefix. It does not take those tokens off the model’s working desk. Calera ICX provides a true AI Memory Substrate and LLM Substrate that keeps the entire knowledge archive in 4D topological lattice memory and streams only a focused, grounded viewport to the LLM.

Without ICX Substrate
Monolithic 1M+ LLM Context Window & Vector RAG
  • Haystack stays on the desk: A cache-read discount (often 10–50% of input) still leaves the prefix in the attention window.
  • New dumps bill at full price: Tool results, grep output, and file reads append every turn. Prefix edits, tool reorder, and naive compaction break the cache.
  • Lost-in-the-middle: Attention still diffuses across the cached haystack. Overflows still fail the call — Gemini-alone missed 188/503 rows in our official run.
Cheaper re-reads • Same haystack • Cache breaks still full-price
With ICX AI Memory Substrate
Decoupled 4D A₄ Lattice Storage Substrate + Scoped Viewport
  • The infinite warehouse: Compile corpora once into 4D A₄ lattice memory on CPU with no ongoing storage fee.
  • The scoped viewport: Streams only 500–1,500 pinpoint Grounded Facts to the LLM per turn.
  • Sub-millisecond recall: <0.5ms lattice traversal with formal boundary refusal (∂² = 0).
Archive stays in lattice • Viewport on the desk • Grounded facts only
Core Substrate Architecture

The Calera Epistemic Tetrad Architecture

Four mathematically unified foundations powering the Calera AI Memory Substrate: geometric routing, continuous topological storage, least-action geodesic navigation, and boundary-grounded hallucination elimination.

Pillar 1 • Geometric Viewport

Fast Geometric Viewport

σ Constant Delay Routing (σ-3 + σ-4 = 1)

Replaces unbounded Softmax quadratic attention with deterministic geometric conservation routing. Delivers up to 88.5% Time-To-First-Token (TTFT) reduction and sustained 128k context windows on standard CPU hardware.

88.5% TTFT Reduction • Zero GPU Starvation
Pillar 2 • Infinite Context

4D Epistemic Coprocessor

Sub-5ms O(1) Lattice Memory (e.mem)

Persistent 4D A4 simplicial complex memory substrate decoupling enterprise knowledge from LLM context limits. Ingests dynamic enterprise facts in real time with continuous learning without backpropagation or weight retraining.

<5ms Lattice Diffusion • 1M+ Node Scaling
Pillar 3 • Geodesic Ribbon 100% Opt-In | Disabled by Default

Epistemic Geodesic Navigator

Stationary Action Path (δS = 0) • 1D Ribbon

Calculates variational shortest paths across the 4D simplicial manifold from premise to conclusion. Serializes verified milestone transitions into an ultra-compact 1D Holographic Ribbon ([nav:ribbon ...]), dropping prompt injection from 2,000+ tokens to ~80 tokens.

<80 Tokens • Optional Acceleration
Pillar 4 • Grounding Firewall

Epistemic Grounding Firewall

Nilpotent Boundary Operator (∂2 = 0)

Formal homological invariant guaranteeing zero statistical hallucination. Grounded knowledge forms closed simplicial cycles (∂c = 0). Non-provable or fabricated claims breach the cycle boundary and are deterministically refused.

0.00% Statistical Confabulation Guarantee

Substrate Architecture Matrix: Baseline Triad vs. Epistemic Tetrad

Geodesic Navigation is an optional high-performance mode. Calera ICX default production endpoints and existing OpenAI-compatible integrations remain completely untouched.

Architectural Dimension Baseline ICX (Default Production) With Geodesic Ribbon (Optional)
Activation Mode Default Active (Standard) 100% Opt-In • Disabled by Default
Memory Representation 4D A4 Topological Memory (e.mem) Continuous 4D Simplicial Manifold with Manifold Lens
Retrieval Output Scoped Grounded Facts (500–1,500 tokens) 1D Holographic Ribbon Sequence (~80 tokens)
Routing Principle Discrete Simplicial Lattice Diffusion (<0.5ms) Variational Geodesic Least Action (δS = 0)
Prompt Injection Format Formatted Grounded Fact Claims Compact Milestone Transitions ([nav:ribbon ...])
Latency / TTFT Advantage Sub-millisecond lattice recall Sub-5ms pathing + additional 40–88% TTFT reduction
Activation Triggers Automatic on all /v1/chat/completions calls Header X-ICX-Geodesic-Ribbon: true, parameter enable_geodesic_ribbon: true, or POST /v1/lattice/geodesic
Interactive Substrate Comparator

Compare Prompt Footprints: Vector RAG vs. Geodesic Ribbon

See why traditional vector retrieval floods the LLM attention window with thousands of noisy tokens, while the Calera 1D Holographic Ribbon serializes only verified milestone transitions.

Traditional Vector RAG / Full Context
2,420 Tokens Injected
Time to First Token (TTFT): ~1,280 ms
Memory Overhead: 14 Unstructured Chunks
Attention Diffusion: Lost-in-the-Middle Risk
Prompt Context Injection Payload:
[DOCUMENT RETRIEVAL RESULT - TOP 14 CHUNKS]
--- CHUNK 1 (AAPL 10-K Item 7, p. 34): "Total net sales were $383,285 million in 2023 compared to $394,328 million in 2022..."
--- CHUNK 2 (AAPL 10-K Item 8, p. 48): "Operating income for the fiscal year ended September 30, 2023 was $114,301 million..."
--- CHUNK 3 (AAPL 10-K Note 5 - Taxes, p. 55): "The provision for income taxes was $16,741 million, reflecting an effective tax rate of 14.7%..."
--- CHUNK 4 (Non-GAAP Reconciliations): "Net operating profit after tax (NOPAT) represents operating income less adjusted income taxes..."
--- CHUNK 5 (Cash Flows): "Cash generated by operating activities was $110,543 million while payments for capital assets were $10,959 million..."
[... 9 additional noisy document passages truncated (2,420 total tokens) ...]
Optional 1D Holographic Ribbon Mode
84 Tokens Injected
Time to First Token (TTFT): 48 ms (-88.5% reduction)
Continuous Geodesic Action (S): 5.66 (Stationary Path)
Grounding Verification: 100% Closed Causal Chain (∂c = 0)
Verified Geodesic Milestones:
AAPL_REV → OPERATING_INCOME → TAX_EXPENSE → NOPAT → FREE_CASH_FLOW
Compact Holographic Ribbon Payload:
[nav:ribbon premise="AAPL_REV" target="FREE_CASH_FLOW" steps=4 action=5.66]
AAPL_REV -> OPERATING_INCOME -> TAX_EXPENSE -> NOPAT -> FREE_CASH_FLOW
# Epistemic invariants verified: ∂² = 0, action S = 5.66, latency = 4.2ms
Enterprise Grade

Features Built for Enterprise Scale

Everything your engineering, security, and product teams need to deploy infinite context memory with zero compliance or performance friction.

Zero-Trust BYOK Security

Bring Your Own Google Gemini or OpenAI API keys. Keys are forwarded ephemerally via request headers (X-LLM-API-Key) with zero server-side key retention or training on customer data.

Drop-In OpenAI SDK Gateway

100% compatible drop-in replacement for standard OpenAI SDKs in Python, Node.js, Go, and Rust. Simply update baseURL and pass X-Space-ID to begin grounding requests.

Hosted Agent MCP Server

Native 1-click Model Context Protocol (MCP) server integration for Google Antigravity, Claude Desktop, Cursor, and custom autonomous agent swarms.

MCP in the docs →

Isolated Memory Workspaces

Multi-tenant namespace isolation (legal_vault, eng_repo, clinical_ehr) ensures zero cross-tenant factual leakage and instantaneous context switching.

Sub-Millisecond Engine Latency

Discrete simplicial lattice diffusion executes on standard CPU memory in <0.5ms, eliminating expensive GPU cluster reservation fees and VRAM memory starvation.

Topological Grounding Invariant

Factual assertions form closed simplicial chains (∂c = 0). Hallucinated or out-of-ontology statements trip the boundary refusal operator instantly.

Zero-Trust BYOK Keys forwarded ephemerally via X-LLM-API-Key with zero server key retention.
Zero Training on Data Customer repositories and document lattices are never used for model training or tuning.
Hallucination Firewall (∂² = 0) Deterministic topological boundary refusal prevents statistical confabulation.
Dedicated Worker Nodes High-throughput Cloud Run & VPC private instances with 1,500 req/min burst SLA.
1-Line Integration

Developer SDK & Agent Quickstart

Connect Claude Desktop, Cursor, LangChain, LlamaIndex, or standard OpenAI SDKs to persistent memory in under 60 seconds.

Want a live comparison before copying code? Open the ICX console preview or watch the SwarmForge multi-agent demo — both use the live ICX lattice memory engine.

pip install calera-agent-memory 
# pip install calera-agent-memory
from calera_agent_memory import CaleraMemoryClient

# Auto-provision a free 1,000,000-node persistent vault (Delaware UETA § 14)
client = CaleraMemoryClient()
vault_info = client.claim_free_vault("developer@company.com", agent_name="my-agent")
print(f"Provisioned Vault ID: {client.vault_id}")

# 1. Store structured facts or task state into 4D topological lattice
client.store("NVDA_FY26_Rev", "Estimated at $168B on Blackwell datacenter expansion")

# 2. Sub-5ms topological associative recall with zero context rot
results = client.query("NVDA revenue projection")
print(results)
W3C WebMCP Standard Ingress

🤖 Invoke ICX With Your AI Agent

Autonomous agents (Claude Computer Use, OpenAI Operator, MultiOn, Browserbase) can claim a dedicated 1,000,000-node persistent memory vault under Delaware UETA § 14 in <300ms.

consequentialHint: true $0.00 Permanent Free Tier
Agent Instructions Prompt 
Navigate to https://icx.caleralabs.com and call the WebMCP tool `calera.icx.claim-1m-nodes` with principalEmail="YOUR_EMAIL_HERE" and acceptA2bTerms=true to self-provision my free 1,000,000-node persistent memory vault.
Paste this prompt into Claude Computer Use, OpenAI Operator, or send via Playwright agent harness.
Live In-Browser WebMCP Test ● modelContext Active 
Click "Execute Tool" to simulate in-browser agent actuation via document.modelContext.
Knowledge & Architecture

Frequently Asked Questions: AI Memory Substrate

Everything you need to know about the AI Memory Substrate, the Substrate Separation Principle, and integrating ICX with your LLMs and agents.

What is an AI Memory Substrate?

An AI Memory Substrate is an external, persistent, mathematically verified cognitive memory architecture that stores factual knowledge, codebases, and conversational history independently of an LLM's transient context window. ICX provides a 4D A₄ topological lattice memory substrate that allows any upstream LLM to recall verified facts in sub-5ms O(1) time without token bloat or attention degradation.

What is an LLM Substrate?

An LLM Substrate is the foundational memory and verification layer that pairs with generative foundation models. By delegating long-term fact storage to the ICX substrate (the Substrate Separation Principle), LLMs avoid quadratic prefill compute costs, eliminate lost-in-the-middle context rot, and maintain persistent state across unbounded multi-turn sessions.

Why does an LLM require an external memory substrate?

Transformers are reasoning engines, not storage databases. In standard monolithic context windows (128k to 1M+ tokens), re-sending long context causes attention diffusion, high latency, and frequent recall failures. An external memory substrate maintains the full knowledge archive in 4D lattice memory on CPU and streams only a laser-focused viewport of 500–1,500 Grounded Facts to the LLM per turn.

How is the ICX Memory Substrate different from Vector Databases and GraphRAG?

Vector databases rely on approximate cosine similarity over dense statistical embeddings, which is non-deterministic and prone to hallucinations. GraphRAG introduces heavy LLM extraction latency and high token costs. ICX compiles corpora directly into discrete 4D A₄ Coxeter simplexes, enabling deterministic simplicial lattice walking in sub-5ms with mathematical boundary refusal (∂² = 0) when information is absent.

Is ICX compatible with OpenAI SDK, Claude, and Cursor?

Yes. ICX is a drop-in replacement for OpenAI SDK endpoints at https://icx.api.caleralabs.com/v1 with full support for /v1/chat/completions, streaming, and tool calls. It also natively supports Model Context Protocol (MCP) at https://icx.caleralabs.com/mcp for instant 1-click addition to Cursor, Claude Desktop, and Google Antigravity.

Can AI agents self-provision an ICX memory vault?

Yes. Autonomous agents can claim a free 1,000,000-node memory vault in <300ms by calling calera.icx.claim-1m-nodes via window.document.modelContext (W3C WebMCP) or POST /api/v1/agent/provision under Delaware UETA § 14 with zero human friction.

Dedicated Infrastructure

Dedicated Enterprise Worker Nodes

Provision isolated, high-throughput Cloud Run worker instances with guaranteed 1,500 req/min burst rates, dedicated CPU/memory specs, and private memory space isolation for enterprise security requirements.