Skip to content

SuperLocalMemory v4.2.12

Choose a tag to compare

@github-actions github-actions released this 28 Oct 06:29
· 1 commit to main since this release
Immutable release. Only release title and notes can be modified.
19e0d82

Fixes for six reports filed after 4.2.11. No migration and no command to run.

Fixed

  • A slow recall answers instead of saying busy (#304). On a store of 37,000
    memories every explicit recall ran past its time limit and came back as HTTP
    503 "ask again in 2 seconds", on an idle daemon, and every retry did the same.
    A recall that runs out of time now answers with keyword matches, labelled
    degraded_lexical and unchecked. "Busy" is kept for when every slot really is
    taken, the one case where asking again helps.
  • Recall does a third of the vector work (#304). Three search channels ran
    the same nearest-neighbour search at the same moment on every recall. They
    now share one, and the channel phase stops waiting before the caller gives up.
  • A purge finishes after fast note changes (#297). Editing or deleting a
    note while its save was still being enriched made that save fail three times
    with "queryable fact profile mismatch or missing fact", and
    sources remove --purge refused until the retries ran out. Such a save now
    settles at once and puts nothing back.
  • A custom embedding endpoint keeps the model you chose (#302). In Modes A
    and B a blank model name was filled with the local model's id and width, which
    an endpoint like Ollama's does not serve, so every embedding call failed.
    Saving such an endpoint without a model name or width is now refused with a
    message saying what to fill in.
  • Modes A and B say when data leaves this computer (#303). A remote
    embedding endpoint, or a hosted language model in Mode B, was accepted while
    the mode still reported local-only. The dashboard card, the mode status, the
    configuration warnings and the EU AI Act check now follow the configuration.
  • Every MCP tool has a title and all four behaviour hints (#300), as the
    Claude and ChatGPT directories require. slm_cache_set no longer claims to be
    read-only.
  • The web sign-in tells identical connections apart (#301). After
    reinstalls the connection picker listed several identical entries. Each now
    shows a short ID and its expiry, and an opt-in box removes the older ones for
    that profile and app. Patch by @DannyGi82.