A persistent semantic graph beneath every agent.
AI coding agents are only as good as the context they hold. Today most of them start every session blind, re-reading the same files to rediscover a structure that never changed. KinLab gives agents a standing semantic graph of your codebase as the source of truth. Your code is a living semantic graph, not files and diffs with AI on top. An agent reads structure instead of re-scanning text.
Every session starts fresh. The agent forgets.
The default agent loop is amnesiac. Open a session, burn a context window re-reading files, do the work, throw the picture away. The next session starts from zero again, same codebase, same structure. Stateless retrieval rebuilds an index it immediately discards, and the agent never accumulates a real understanding of the code.
Re-reading the same files
Context window spent rediscovering structure that has not changed since the last session. A tax paid on every single run.
Statelessness forgets
Per-session retrieval has no memory. What the agent learned about your code last time is gone the moment the session ends.
No shared picture
Each agent builds its own private, partial view. Nothing connects them to a single, current understanding of the codebase.
The graph is the source of truth
KinLab does not hand agents a fresh index every session. It gives them a persistent semantic graph: a standing model of your code's structure that stays current and is shared across every agent and human on the project.
A standing semantic graph
Your codebase is parsed into a persistent graph of entities (functions, types, contracts) and the relationships between them. It is the source of truth that stays current, not a throwaway index an agent rebuilds and discards every session.
MCP-native, every agent
A Model Context Protocol server exposes semantic tools (locate, context packs, trace) to any MCP client. One command, kin setup, wires Claude, Cursor, Codex, and Gemini onto the same graph, so context is not a per-tool integration project.
Structural context, not snippets
When an agent asks about a symbol, it gets the structure: what calls it, what it depends on, the contracts it implements, the tests that exercise it. Context arrives as relationships an agent can reason over, not a bag of loosely-related text.
Wire it into your agents
$ kin setupOne command auto-wires the MCP server into Claude, Cursor, Codex, and Gemini. The agents you already use get graph-native retrieval, no bespoke integration per tool.
Versus per-session indexes and memory tools
A throwaway index rebuilt per run, or a bolt-on memory store of past chats, both patch around the same gap. A graph closes it, because it is persistent, writable, shared, and governed.
Persistent. The graph outlives the session instead of being rebuilt from scratch each time.
Writable. Agents change the graph through governed tools, not just read a static snapshot.
Shared. Every agent and human works against one source of truth, not a private per-agent cache.
Governed. Access is bounded and every change carries provenance, so shared context stays auditable.
Persistent, not per-run
The graph stands between sessions. Agents pick up structural context that is already current instead of rebuilding it from zero.
Shared and writable
One graph for the whole team and fleet. Agents read and change it through governed tools, so context is common ground, not a private cache.
Governed by default
Shared context only works if it is bounded. Access is scoped and every change carries provenance, so the common graph stays trustworthy.
Pre-release · early access by request
Stop re-reading files. Stand a graph beneath your agents.
KinLab is the AI-native source-control and collaboration platform on the open Kin substrate. Early access is granted by request while the platform matures.
Read the kin-mcp docs, browse the Kin ecosystem, or see the reproducible proof.