SuperLocalMemory 4.1.18 — Recall can say "I don't have that"

[4.1.18] — Recall can say "I don't have that"

Added

• Answer check. Recall now says whether the memories it found actually
  answer the question, instead of handing back the best of a bad set. Results
  are never removed or hidden: each recall reports abstained, the reason,
  a confidence, and answer_check_status (judged, off, skipped, busy,
  warming or unavailable), and every connected agent is told not to present
  memories that do not answer as the answer.
• Two ways to run it, one at a time. "On this Mac" runs privately on Apple
  Silicon and nothing leaves the machine. "Online with Jev" works anywhere with
  your own TypeSafe or OpenRouter key; it is off until you choose it and tick a
  consent box, and it sends only the question and its top three memories,
  with credentials stripped out first. It never sends another profile's
  memories, and recalls that only load context (session start, auto-loaded
  context, the per-prompt hook) are never sent. Turning one on turns the
  other off.
• Optional: let Jev reorder results. Inside "Online with Jev" there is a
  separate switch, off by default with its own notice, that also asks Jev
  which of the top results answers the question most directly and returns
  them in that order. Still one request per recall. On a public benchmark of
  long conversations (LoCoMo, 1,531 questions) the memory that answers came
  first 62% of the time instead of 49%; it adds about 0.4 seconds to a
  recall. If the provider is slow or unreachable, results keep their usual
  order.
• Set up from the dashboard. Settings → Answer check: choose, set up with a
  progress bar, use an existing install on networks that block downloads,
  save and test a key (only its last four characters are ever shown), or turn
  it off. The dashboard home shows what is actually running. Changing it
  needs the manage permission, and withdrawing consent stops the online
  check at once, in every mode.
• First-time setup installs it on Apple Silicon (about 1.1 GB to
  download, about 1.1 GB of memory while on). If the download is blocked or
  interrupted, setup finishes and Set up again picks up where it left off.
  Unattended setup skips it unless SLM_INSTALL_LAYA=1 is set.
• Every recall response says what chose its order and what became of
  the answer check, from the MCP tools, the CLI and the HTTP API alike.

Fixed

• Archived memories stay archived after a restart. A repair that should
  have run once ran on every start and put archived memories back into
  recall.
• Upgrades keep one safety copy instead of two, and copies left by
  earlier upgrades are cleaned up.
• Ranking no longer loses its second pass. After a crash, a restart of its
  helper, or an idle shutdown, the reranking step recovers on its own, and it
  keeps ranking while its helper restarts. It also uses much less memory on
  long, mixed-length requests.
• A configured remote reranker is now used, and one that stops answering
  no longer slows every recall: recall falls back to its usual ranking right
  away and checks the endpoint again in the background.
• A newer memory is no longer pushed down by an older one about the same
  subject.
• Health checks report the reranking step and no longer raise false alarms
  when a working search simply has no match.
• No more project files at the root of the disk. Background context sync
  only writes into real project folders.
• Bounded loops waiting on memory no longer finish on a recall that does
  not answer their question.
• Restoring a backup no longer loses or corrupts recent memories, takes
  effect while the store is in use, and no longer leaves a large temporary
  file next to the store.
• Stored memories keep their links and file paths. Saving a memory no
  longer blanks out long URLs and paths.
• slm db compact works while SLM is running.
• Hooks work when the data folder has a long path.
• Maintenance no longer reports a step that does not apply to your store
  as a failure.
• Windows: safety copies are flushed to disk properly.

Security

• A web page on another site can no longer reach SLM through your
  browser. SLM now answers only requests addressed to this computer. To
  reach a LAN install by a host name, list the name in SLM_ALLOWED_HOSTS;
  reaching it by address, or by a name already allowed for network access,
  is unchanged.
• The per-prompt hook only talks to your own SLM service. Its fallback
  path could send a prompt to any program listening on the service's port.
• Known API key formats, webhook links and private-key bodies are removed
  when a memory is saved, not only before anything is sent online.
• Dependencies updated to versions that fix published security advisories.

Changed

• A background consistency check is off by default, including on upgraded
  installs. It changed no recall result in testing and ran on every save. It
  can be turned back on, and turning it back on now sticks.
• Answer-check settings live in one file for every mode, so switching
  modes never changes them.
• Mode A says when it is not local-only: with the online answer check on,
  the mode description says so instead of "zero-cloud".

Known limits

• Macs shared by several user accounts: another account on the same Mac
  can reach the local SLM service. Turn on company mode (Team → Access
  policy) on a shared Mac; a per-account check is planned for the next
  release. See the "Shared Macs" section of the security notes.