Introduction
Introduction
Section titled “Introduction”AgentCTX is a cognition layer for AI agents. It gives every agent — Claude, Gemini, GPT, Llama, or whatever comes next — structured memory, scoped tool access, semantic knowledge retrieval, and cryptographically verifiable trust, all through a single gateway and a purpose-built query language called CTX.
The Problem
Section titled “The Problem”Every session, your agent starts empty. No memory of yesterday. No recall of the tools it used, the bugs it fixed, or the architecture it debated for hours. It re-reads files, re-discovers tool schemas, and reconstructs understanding from scratch.
This costs you tokens, time, and quality:
| Waste Category | Without AgentCTX | With AgentCTX | Savings |
|---|---|---|---|
| Tool descriptions (70 tools) | 9,950 tokens | 850 tokens | 91% |
| Tool invocations (150 calls) | 7,350 tokens | 1,800 tokens | 76% |
| File search + read | 16,925 tokens | 1,290 tokens | 92% |
| Swarm coordination | 21,000 tokens | 860 tokens | 96% |
| Cross-session resumption | 6,500 tokens | 214 tokens | 97% |
| Total per session | 96,645 tokens | 8,161 tokens | 91.6% |
That freed context window isn’t just cheaper — it lets your agent think. More room for reasoning, fewer hallucinations, longer autonomous runs.
How It Works
Section titled “How It Works”AgentCTX introduces CTX — the first query language designed for how agents actually process information. Instead of verbose JSON-RPC or natural language, agents write structured operations:
?k "auth" #code @7d ^3 search knowledge (9 tokens, was 46)>t github.issues.create title="Fix SSE" call a tool (25 tokens, was 67)+m auth-decision #arch "JWT for API" store a memory (12 tokens, was 49)?s "process assets" ^3 recommend skills (8 tokens)Every operation flows through one gateway and targets one of seven planes:
| Plane | Symbol | Purpose |
|---|---|---|
| Tools | t | Discover, inspect, and call MCP tools |
| Knowledge | k | Semantic search across ingested documents |
| Memory | m | 5-layer persistent memory with decay and consensus |
| Skills | s | Reusable workflows and agent capabilities |
| Agents | a | Multi-agent coordination and delegation |
| Inspection | i | Runtime introspection of the CTX system |
| LLM | l | Direct language model inference and embedding |
Trust Without Compromise
Section titled “Trust Without Compromise”Every action an agent takes through AgentCTX produces a signed audit trail. The sidecar translator converts structured CTX into human-readable markdown, signs it with Ed25519, and logs it. No LLM in the translation path. No hallucination risk. Run actx verify and the math settles any dispute.
Agent writes: +m "decision" #arch "JWT for API, session tokens for WebSocket"Human reads: "Stored architecture decision: JWT for API, session tokens for WebSocket"Signature: Ed25519 signed — verifiable with `actx verify`Who It’s For
Section titled “Who It’s For”- Agent developers building tools, frameworks, or autonomous systems
- Platform teams operating multi-agent deployments at scale
- Enterprise teams that need verifiable audit trails for AI operations
- Anyone tired of watching agents waste 90%+ of their context window on tool schemas they’ve already seen
Next Steps
Section titled “Next Steps”- Quick Start → — install, configure, and run your first CTX operation in 5 minutes
- CTX Language → — learn the seven operators and seven planes
- Architecture → — understand the gateway, sidecar, and memory layers