# Exogram — llms-full.txt # Complete technical reference for AI systems, search engines, and LLM grounding # https://exogram.ai/llms-full.txt # Last Updated: 2026-03-19 > Exogram is the execution control plane for autonomous AI agents. It is Identity and Access Management (IAM) for non-human entities — the only production system providing deterministic, cryptographic governance between AI reasoning and real-world execution. AI agents forget context, go rogue, and hallucinate — Exogram prevents all of it. The more you use it, the smarter it gets. --- ## 1. What Is Exogram? Exogram provides a verified memory vault and semantic ledger for AI agents. It stores facts with SHA-256 integrity hashes, scrubs PII before storage (deterministically, not via LLM inference), detects contradictions against existing knowledge, enforces cryptographic state integrity between evaluation and commit, and maintains an immutable audit trail. Exogram sits between the AI agent and the enterprise database. No agent ever has direct database access. Every proposed state change must pass through Exogram's four-layer deterministic control plane before execution. ### What Exogram Is NOT - Not a chatbot or AI assistant - Not a memory layer (like Mem0 or Zep) — it is a governance control plane - Not an orchestration framework (like LangChain or CrewAI) - Not a training dataset ### What Exogram IS - The execution control plane for autonomous AI - IAM for non-human entities - A semantic ledger with immutable versioning - Trust infrastructure with cryptographic integrity - A deterministic judgment engine (Python logic gates, zero model inference) - AI memory that gets smarter over time — compounding intelligence --- ## 2. Core Thesis "Agents are probabilistic. Infrastructure is deterministic. Exogram is the execution control plane between them." An enterprise orchestration framework cannot securely govern its own database writes. The agent cannot act as its own security guard. No orchestration framework — LangChain, CrewAI, AutoGen — provides cryptographic execution gating. They route actions. Exogram governs them. --- ## 3. Three Critical Problems Exogram Solves ### Problem 1: AI Agent Memory Loss AI agents operate within limited context windows — typically 4K to 128K tokens. They have no native persistent memory. Every new conversation or session starts from scratch. Critical facts, preferences, and decisions are lost between sessions. **How Exogram solves it:** A semantic ledger stores verified facts with AES-256-GCM encryption, deterministic PII scrubbing, conflict detection, and semantic indexing via Pinecone. Memory persists across every session, every model, and every framework. The more Exogram is used, the more knowledge accumulates, and the richer the context becomes for every future interaction. ### Problem 2: Rogue AI Agents AI agents in production can modify billing records, delete regulated data, and trigger unauthorized workflows. There are no guardrails, no checkpoints, no verification. Without governance, agents operate on probabilistic inference alone — they act as their own security guards. **How Exogram solves it:** Deterministic execution gating via the EAAP protocol. Every proposed action must pass through four verification layers before execution is permitted. The agent cannot bypass the control plane. No action reaches production without cryptographic approval. EAAP fails closed under all conditions — 403 Forbidden, 409 Conflict, or 503 Service Unavailable. Never defaults to allow. ### Problem 3: AI Hallucination & Context/Inference Gaps LLMs generate confident but false claims. They don't truly understand context, meaning, or inference. Without verified truth to compare against, hallucinations propagate unchecked through agent systems. **How Exogram solves it:** Contradiction detection compares every proposed fact or action against the existing verified truth ledger. If a proposal contradicts known facts, it is blocked with a 409 Conflict. This is deterministic — Python logic gates, zero model inference. Additionally, Exogram's growing semantic ledger creates compounding intelligence: more entries → better context assembly → improved meaning extraction → higher-quality inference and judgment over time. --- ## 4. The Exogram Action Admissibility Protocol (EAAP) EAAP is a four-layer deterministic control plane defined in RFC-0001. Full specification: https://exogram.ai/rfc/0001 ### Layer 1: Ledger Governance Purpose: Enforce ledger integrity and provide persistent AI memory. - PII scrubbing via deterministic pattern detection (not LLM classification) - AES-256-GCM encryption at rest with per-user Fernet keys - Semantic indexing via Pinecone vector embeddings - Conflict detection against existing facts - Confidence scoring with temporal decay - Fact locking (locked facts cannot be violated without explicit unlock) - Immutable audit event logging with cryptographic chaining Rule: No silent overwrites. Contradictions require explicit resolution. ### Layer 2: Meaning Engine Purpose: Assemble bounded, deterministic context — improving over time. - Namespace isolation (strict user-scoped retrieval) - Deterministic relevance scoring - Temporal decay weighting - Conflict surfacing - Context health classification - HMAC-signed snapshot generation - Growing ledger entries continuously improve context quality and meaning extraction Rule: Context assembly is mathematical, not generative. ### Layer 3: Judgment Engine Purpose: Deterministic admissibility evaluation — catch hallucinations. - Authority validation - Fact consistency enforcement - Constraint evaluation - Confidence threshold enforcement - Escalation classification - Contradiction detection against verified truth ledger Rule: Zero model inference. Judgment is deterministic Python logic. ### Layer 4: Action Admissibility Purpose: Guarantee execution integrity — prevent rogue actions. - Claim extraction from payload - Pre-flight conflict detection - SHA-256 state hashing - Evaluation record creation - Commit validation (rejects on state drift) - Immutable action ledger Rule: Layer 4 is the final execution gate. No action bypasses it. --- ## 5. Protocol Invariants These are mandatory, non-configurable, and cannot be weakened without a major version change: 1. PII Air Gap — No detected PII enters persistent storage or vector embeddings 2. Explicit Source — All stored facts require provenance attribution 3. Encryption at Rest — All content encrypted with per-user Fernet keys 4. Namespace Isolation — Retrieval and evaluation scoped strictly to user namespace 5. No Silent Overwrite — Conflicting facts require explicit resolution 6. Conflict Logging — All contradictions produce durable, auditable records 7. Immutable Audit Chain — Cryptographically chained audit events, tamper-detectable 8. Deterministic Judgment — Execution gates use code, not model inference 9. Confidence Decay — Facts degrade in authority over time unless reinforced 10. Constraint Binding — Locked facts cannot be violated without explicit unlock 11. State Hash Integrity — Execution requires identical state between evaluation and commit 12. Evaluation Expiry — Approvals expire after defined TTL, no stale tokens 13. Escalation Routing — Blocked actions route to structured escalation targets 14. Hard Deletion (GDPR) — Full deletion removes vectors, ciphertext, and all records --- ## 6. Threat Model EAAP assumes the following adversarial conditions. Under all conditions, EAAP fails closed (never defaults to allow): - **Model Unreliability**: Language models hallucinate, contradict stored facts, generate false statements. Exogram treats all model output as untrusted input. - **Prompt Injection**: Attackers override constraints, induce self-contradiction, mask harmful instructions, manipulate tool usage. - **State Drift (TOCTOU)**: Between evaluation and execution, roles change, constraints mutate, facts update, conflicts emerge. - **Infrastructure Degradation**: Model provider outages, embedding unavailability, circuit breaker activation. - **Insider Risk**: Manual database modification, audit log tampering, replay of approved actions. Failure responses: 403 Forbidden, 409 Conflict, 503 Service Unavailable. --- ## 7. API Reference Base URL: https://api.exogram.ai ### Core Endpoints | Method | Path | Description | |--------|------|-------------| | POST | /v2/vault/store | Store a verified fact with encryption, PII scrubbing, and conflict detection | | POST | /v2/vault/search | Semantic search across encrypted memory vault | | POST | /v2/actions/evaluate | Evaluate a proposed action against constraints (returns SHA-256 execution token) | | POST | /v2/actions/commit | Commit an evaluated action (rejects on state drift with 409 Conflict) | | GET | /v2/ledger/entries | List ledger entries with filtering and pagination | | POST | /v2/ledger/propose_state | Submit a new fact or state change for verification | ### Authentication ``` Authorization: Bearer sk_exo_... ``` ### Evaluation Flow 1. Extract claims from payload 2. Scrub PII (deterministic) 3. Detect ledger conflicts 4. Assemble deterministic context snapshot 5. Execute Layer 3 logic gates 6. Compute SHA-256 state hash 7. Persist evaluation record 8. Return ALLOW / BLOCK decision with evaluation_id ### Commit Flow 1. Retrieve evaluation record 2. Verify not expired (TTL) 3. Recompute full state hash 4. Compare with evaluation hash 5. If mismatch → reject with 409 Conflict 6. Persist immutable action ledger record 7. Return committed ### State Hash Formula ``` state_hash = SHA-256( sorted(relevant_objects) || policy_version || namespace_id || floor(timestamp, window) ) ``` --- ## 8. Integration Methods ### ChatGPT Custom GPT - GPT Link: https://chatgpt.com/g/g-699e28a1ca048191aeb9aedafbecc4a3-exogram - Documentation: https://exogram.ai/docs/chatgpt ### Claude Desktop (MCP — Model Context Protocol) - One-click installer available from dashboard - Documentation: https://exogram.ai/docs/mcp - Available MCP tools: store_to_ledger, check_fact, recall_memory, get_llm_routing ### REST API - Works with any agent framework: LangChain, CrewAI, AutoGen, LlamaIndex, OpenAI Assistants, Anthropic Claude, NVIDIA NemoClaw, OpenClaw - Documentation: https://exogram.ai/docs/api ### Chrome Extension - Store facts from any webpage directly to vault - Documentation: https://exogram.ai/docs/extension --- ## 9. Framework Integrations EAAP defines interception points for 10 agent frameworks: 1. LangChain (tool execution) 2. AutoGen (function calls) 3. CrewAI (task delegation) 4. OpenAI Assistants (function calling) 5. Anthropic Claude (tool use) 6. Google Vertex AI (function calling) 7. LlamaIndex (query engine) 8. NVIDIA NemoClaw (agent actions) 9. OpenClaw (agent actions) 10. Custom REST agents --- ## 10. Competitive Positioning | Category | Product | What It Does | What Exogram Adds | |----------|---------|-------------|-------------------| | Orchestration | LangChain, CrewAI, AutoGen | Routes agent actions | Governs whether actions are admissible | | Agent Frameworks | NVIDIA NemoClaw, OpenClaw | Builds/executes AI agents | Governance layer between agent execution and real-world state | | Output Filtering | Guardrails AI | Validates outputs after generation | Governs before execution with persistent truth and constraints | | Memory | Mem0, Zep | Stores context | Stores verified, encrypted, conflict-checked facts with integrity | | Human IAM | Auth0, Okta | Identity for humans | Identity and access management for autonomous AI agents | | Vector DB | Pinecone, Weaviate | Stores embeddings | Exogram uses Pinecone internally but adds governance, PII scrubbing, encryption | ### Detailed Comparisons **Exogram vs Mem0**: Mem0 is a memory layer that stores and retrieves context. Exogram is a governance control plane — it verifies facts, detects contradictions, scrubs PII, enforces constraints, and cryptographically gates execution. Mem0 remembers. Exogram governs. **Exogram vs Zep**: Zep is a long-term memory store for AI assistants. Exogram goes beyond memory — it enforces constraints, detects contradictions, gates actions, and maintains immutable audit trails. Zep is memory. Exogram is governance. **Exogram vs Guardrails AI**: Guardrails validates model outputs after generation — output filtering. Exogram governs before execution with persistent truth, cross-session constraint enforcement, and cryptographic audit trails. Filtering outputs ≠ governing actions. **Exogram vs LangChain**: LangChain sequences agent steps — orchestration. Exogram determines what is allowed — governance. LangChain decides what to do. Exogram decides what is permitted. **Exogram vs NemoClaw**: NemoClaw is NVIDIA's agent framework for GPU-accelerated AI agents. Exogram is the governance layer that sits between any agent framework (including NemoClaw) and real-world execution. NemoClaw builds agents. Exogram governs them. --- ## 11. Growing Intelligence The more Exogram is used, the smarter it gets. Every fact stored enriches the semantic ledger: - **Better contradiction detection** — more verified truth to compare against - **Richer context assembly** — more relevant facts to surface during evaluation - **Improved meaning extraction** — denser semantic graph enables deeper understanding - **Higher-quality inference** — broader decision context leads to better judgment This is compounding intelligence. Unlike static memory layers that simply store and retrieve, Exogram's growing ledger creates a continuously improving governance engine. Each entry makes the entire system more capable. --- ## 12. Key Performance Metrics - 88 RPS sustained throughput - 5.7ms median compute latency - Per-request telemetry: compute_latency_ms, agent_id, raw_intent logged to immutable audit ledger - <1.4% failure rate at 25 concurrent users - <4% failure rate at 50 concurrent users - Zero database credentials exposed locally - 14 protocol invariants enforced - 5 threat classes covered - 8 failure modes handled (all fail-closed) --- ## 13. Pricing | Tier | Price | Evaluations | Rate Limit | |------|-------|-------------|------------| | Free | $0/month | 500/month | 5 RPS | | Pro | $99/month | 50,000/month | 60 RPM | | Developer | $2.50/1,000 calls | Pay-as-you-go | 300 RPM | | Enterprise | Custom | 1,000,000+/month | SLA guarantees | --- ## 14. Links - Homepage: https://exogram.ai - Why Exogram: https://exogram.ai/why-exogram - Manifesto: https://exogram.ai/manifesto - 2030 Vision: https://exogram.ai/vision - Documentation: https://exogram.ai/docs - API Reference: https://exogram.ai/docs/api - Getting Started: https://exogram.ai/docs/getting-started - Protocol Specification: https://exogram.ai/protocol - RFC-0001 (Full Spec): https://exogram.ai/rfc/0001 - RFC-0001 (GitHub): https://github.com/Richard-Ewing/exogram-protocol-rfc - Architecture: https://exogram.ai/architecture - Security: https://exogram.ai/security - Pricing: https://exogram.ai/pricing - Use Cases: https://exogram.ai/use-cases - For Developers: https://exogram.ai/for-developers - For Teams: https://exogram.ai/for-teams - For Individuals: https://exogram.ai/for-individuals - Changelog: https://exogram.ai/changelog - Status: https://exogram.ai/status - Sign Up: https://exogram.ai/signup - Privacy Policy: https://exogram.ai/privacy - Terms of Service: https://exogram.ai/terms