SuperLocalMemory v4.2.9

If you run Mode B with Ollama, every saved memory used to cost up to eleven
model calls, most of them wasted. A machine with a steady stream of memories
kept Ollama busy all day, and recalls waited behind it. Nothing to run after
upgrading.

Fixed

• Linked memories no longer trigger model calls that were never saved.
  Each new memory links to up to ten similar older ones, and Mode B asked the
  model to rewrite every one of them. The request never included the new
  memory, and saving the answer always failed without a visible error, so none
  of that work was ever kept (since 3.2.2). Linked memories now get the new
  memory's keyword added, as in Mode A, with no model call.
• Thinking models answer directly. SLM never told Ollama to skip a model's
  thinking step, so models that think by default (Gemma 4, Qwen 3) reasoned
  for ~400 tokens before every short answer: about 5 s instead of under 1 s on
  gemma4:e2b, and an empty answer when the reply had a length limit. Thinking
  is now off for every SLM request. Set SLM_OLLAMA_ALLOW_THINK=1 to give the
  model its default back.
• Memory descriptions from Gemma 4 are kept. Gemma 4 wraps its answer in a
  code fence, which SLM could not read, so the description fell back to a
  template every time. Fenced answers are now read.
• slm provider set ollama --mode b --model <name> does what it says.
  Named providers were always saved as Mode C with the preset model: --mode
  and --model were read only for custom. A local provider now defaults to
  Mode B, --mode and --model are honoured, and an unknown mode is refused.
• A model or embedder you change stays changed when you switch modes.
  SLM keeps one settings file per mode. Running slm mode b while already in
  Mode B, or switching modes from the dashboard, skipped saving the mode you
  were in, so the next switch could bring back the old model or embedder. The
  mode you leave is now always saved first, from the CLI and the dashboard.
• Setup recommends the model that tested best. An installed gemma4:e2b
  was put first in the model list and offered as a 7.2 GB download, although
  SLM has not tested it. The list now puts models that kept facts intact in
  SLM's extraction test first, then untested ones, then models the test caught
  inventing details (such as llama3.2).

Measured on one Mac with gemma4:e2b: a memory's description took 11.7 s and
was discarded; it now takes 2.2 s and is kept. The calls for three linked
memories took 28.7 s; they now take none.