SuperLocalMemory 4.2.6

No migration. Nothing is moved or deleted.

Hermes

• The plugin did nothing on most installs. Hermes refuses a plugin's MCP
  calls until config.yaml grants the server under
  plugins.entries.superlocalmemory.mcp_allowlist, and it does not ask for
  that grant; the install guide said it would. The plugin treated the refusal
  like any failure, so recall, session tracking, tool logging and turn capture
  were silently off, and slm upgrade-hosts reported Hermes as verified. Now:
  • slm upgrade-hosts --host hermes --apply writes the grant (a backup is
    kept; a config with hand-written comments is never rewritten, you are shown
    the line instead);
  • slm doctor has a "Hermes plugin access" check;
  • the plugin warns once in Hermes's log with the fix, and /slm status leads
    with it;
  • the guide shows the step.
• capture_turns: true kept almost nothing (#263). Turns went to the
  auto-capture path, which keeps only English sentences about a decision, a bug
  or a preference. Every completed turn is now saved as a memory, in any
  language, with secrets and email addresses redacted first; a retried turn is
  saved once.

Checked on a real Hermes: its hooks (tool logging, turn capture, recall from an
earlier session), 16 skills, the /slm commands and 4 agent tools.

Plugin directories

• Claude Code: the plugin passes claude plugin validate --strict. It now
  has a README stating what each hook does and where data goes, its licence
  file, and licence and homepage in the manifest. The agent rules no longer sit
  in a root CLAUDE.md that Claude Code never loaded; the skills teach them.
• Cursor and Grok Bot: the Cursor-format plugin is its own folder,
  cursor-plugin/, under the MIT licence (Cursor's marketplace does not accept
  copyleft). It is a launcher for the separately installed engine, which stays
  AGPL-3.0-or-later.
• VS Code and Antigravity IDE extension: a new MIT extension,
  vscode-extension/, adds SuperLocalMemory's MCP server to the editor
  (VS Code 1.101 or later) when slm is installed. It adds nothing when that
  editor's own MCP config already has SuperLocalMemory (VS Code's mcp.json,
  Antigravity's mcp_config.json), so there is never a second copy. Cursor
  does not support the VS Code MCP provider API the extension uses; in Cursor
  use the Cursor plugin or slm connect cursor. The .vsix is attached to the GitHub release. (A first build
  left one of its two modules out of the package, so it failed to start in
  VS Code; a test now checks every module it loads is packaged.)
• Codex: the plugin carries its icon, logo, privacy and terms links.
• Official MCP Registry: server.json ships in the repository and moves
  with every release.

One tool set for every app

• slm tools shows which MCP tool set your apps get and how many tools
  each set has; slm tools set <name> changes it for every app at once
  (slm tools set default goes back). Restart the apps to see it. An app whose
  own config sets SLM_MCP_PROFILE keeps that value.
• The default stays full, 58 tools: all of memory, SLM-Mesh, optimize and
  loops, under the roughly 100-tool limit some hosts put on all MCP servers
  together. No plugin or extension forces a set any more: the Cursor and Grok
  Bot plugin used to give 19 tools and 6 skills (now the default and all 16
  skills), the Antigravity plugin 70.
• The code-graph and governance skills now tell the agent what to say when
  their tools are not in the current set (slm tools set code or power).

New hosts

• Grok CLI: slm connect grok-cli writes the MCP server into
  ~/.grok/config.toml, and Grok CLI runs the Claude Code plugin as is
  (grok plugin install <path>/plugin --trust): 16 skills, 4 agents and the
  hooks. One hook did not start there, because Grok runs a command that is only
  a quoted path as a file name with the quotes kept; every hook command now
  starts in Grok CLI and Claude Code alike. Grok sends hook data in camelCase
  (sessionId, toolName); every slm hook now reads both spellings, so tool
  outcomes are recorded there too. slm connect grok-cli honours $GROK_HOME.

• Muse CLI: slm connect muse writes the MCP server into Muse's
  settings.json and keeps every other setting. A file that still has the old
  mcp_servers key gets it renamed to mcpServers in the same write, since
  Muse ignores every server when both are present. A new muse-plugin/ brings
  the 16 skills, 4 agents, /slm-loop and the hooks: the Claude Code plugin
  without the allowed-tools lines Muse rejects and without an MCP launcher,
  and with a session-start hook that answers in JSON (Muse drops output that
  starts with [, which the plain-text one does). Its MCP entry is marked
  optional, so Muse never aborts a run because SuperLocalMemory did not start,
  and its agents name the tools the way Muse does (mcp__superlocalmemory__…).

• VS Code's Agents window and GitHub Copilot CLI: slm connect copilot-cli
  writes ~/.copilot/mcp-config.json, which both read (the Agents window does
  not see servers from editor extensions). vscode-agent-plugin/, listed by
  .github/plugin/marketplace.json where VS Code looks first, brings the skills
  and agents with a plain slm mcp server: the Claude Code plugin's launcher
  uses a shell default VS Code cannot run.

• OpenClaw: slm connect openclaw adds the server under mcp.servers in
  openclaw.json and copies 14 skills into ~/.agents/skills, where OpenClaw
  loads them. A JSON5 file with comments is never rewritten; the command to run
  is printed instead.

Packaging

• The npm package would have shipped no Cursor plugin. It moved to its own
  cursor-plugin/ folder in this release, and package.json still listed only
  the old folders. Both cursor-plugin/ and .cursor-plugin/ ship now, and a
  test checks every plugin folder the guide points at.

Fixes

• Rolling back the text model could start the picture model's worker
  (#262). A previous model saved under the picture worker's provider was loaded
  through that worker, which already held its own model; on a small machine the
  rollback stopped for lack of memory. Rollback now loads it under the automatic
  provider. The CLI and the dashboard share this path.