The application is the memory.
Coding agents forget what they worked out. Cratonis serves app platforms whose own database is the agent’s memory — so any change to the application is a change to the memory, and nothing has to be explained twice.
19 filed inventions Bring your own agent Multi-tenant SaaS
module billing.invoice rev 41 schema tenants.plan_tier changed 2m ago decision soft-delete over cascade agent note retry window is 90s, not 30s endpoint /brief?scope=billing
Three properties no prompt file gives you
Compiled, not retrieved
Typed SQL over eight typed sources, in a defined sequence. No embeddings, no vector store, no similarity search — the same database state produces the same document, byte for byte, every time.
One artifact, every tool
The compiled document is served live over MCP and written into the instruction files your agents already read — CLAUDE.md, AGENTS.md and five more. Same bytes both ways, with a digest so a stale file is detectable.
Your agents, your keys
Cratonis runs no model of its own. Claude Code, Cursor, Copilot, Gemini — your agents connect in and bring their own accounts, so your context layer never becomes an AI bill.
Memory that closes across sessions
-
Compile
Your application's own Postgres — records, identity, stack, modules, history, custom guidance — compiled into one instruction document, visibility-filtered and priority-sorted.
-
Deliver
Agents read it over an authenticated MCP endpoint, or from the version-controlled files in your repository when the network is not there. Both carry the same compilation.
-
Write back
What an agent learns, it records — through a validated write-back tool that stamps AI authorship, so a guess can never masquerade as a human decision in a later session.
-
Recompile
Mutation observers watch every source. A write triggers recompilation without anyone asking, and the next session — any agent, any tool — starts from the updated document.
What the substrate actually does
Every claim here describes a shipped mechanism, not a roadmap. Where something is scheduled rather than shipped, it says so.
Eight typed sources
Each compiled by typed SQL in a defined sequence, with hard byte budgets so one source can never crowd out the rest.
Scoped, prioritised context
A critical decision outranks a low-priority note deterministically, not by embedding distance.
Hosted MCP endpoint
Agents connect over Streamable HTTP with a bearer credential, plus focused read tools for search, pages and history.
Seven instruction files
Spliced into the files your tools already read, inside managed markers that never touch your own notes.
AI authorship, recorded
Agents persist what they learn through a validated tool that forces origin to ai — an agent cannot label its own guess as a human decision, and every later session can tell the two apart.
A gated write path
File writes pass seven ordered gates, land in staging with an append-only ledger entry, and stay staged until a human promotes them.
Role-based access control
Per-user denial overrides beat any granting role — with a super-admin that no denial can lock out of the room.
Your database, your context
One application and one Postgres database per tenant. Delete the platform; your app keeps working.
Modules tier
Not yet shipped. A dashed frame marks anything on the schedule rather than in the product.
Built for the corporate buyer, not the demo
Role-based access with per-user overrides. An append-only ledger of what every agent did. A database-per-tenant architecture where your context never leaves your own application — delete our platform and your app keeps working.