Each memory stays mapped to the files, functions, and lines it concerns, so retrieval is keyed to code rather than to a conversation.

memkith integrate writes native setup for Claude Code, Cursor, Codex, and Zed.

Each memory records the branch and commit its code lives on. Today that is an address, not a filter: recall is scoped to the repository, and branch-scoped recall sits under Next.

The team boundary. Memory inherits the access rules the code already has, with no second permission model to keep in sync.

memkith resolve stamps the commit onto memories learned before their code landed, and install-git-hook runs it automatically after each commit.

memkith login --no-browser prints a URL instead of opening one, for terminals that cannot.

Installation and repository permissions are handled through a GitHub App rather than through a personal token.

Recall filtered to the branch you are on, with a memory promoted when its branch merges. The branch is already recorded on every memory; the filtering is not built.

The agent list is expected to grow. Any MCP-capable agent can already be pointed at the server by hand.

The memory model, the MCP server's tool surface, and team permissions are the three areas with nothing written.

Neither exists. No audit has begun and no business associate agreement is signed today. The mechanisms that do run are listed on pricing.