SuperLocalMemory 4.1.19 — Memories know what kind they are

Added

• Nine memory kinds. A memory can be a fact, an event, a current status, an
  opinion, a standing rule, a decision, a how-to, a plan or a correction. You
  or your agent can say which when saving (kind on remember, --kind on
  the CLI); otherwise SLM suggests one. A suggestion changes nothing until
  someone confirms it.
• Standing rules at session start. Rules and decisions you confirmed are
  given to every new agent session, up to ten rules and five decisions.
  Proposed, rejected, archived, quarantined and replaced items are never
  included.
• Classify existing memories in the background (slm kinds backfill, or
  HTTP), with progress, pause, cancel and an exact undo. It never changes a
  kind you confirmed. Classifying online with Jev needs its own consent and a
  confirmation that memory text leaves this machine.
• Review and confirm kinds from agents (four new tools in the default,
  code, full and power tool sets), the CLI (slm kinds set | review | confirm)
  and HTTP.
• Ask for a kind of memory. Recall, search and list take a kind filter.
  "What did we decide" favours decisions, "how do I" favours how-tos, and "what
  is the current state" puts the newest confirmed status or decision first.
  Questions that ask for no particular kind are ranked exactly as before.
• Narrow a recall by project, by the agent that saved a memory, or by who
  or what it is about (project, saved_by, about). The filter applies on
  every recall path and before the answer check, so the check judges only what
  you will see.
• Say which memory a new one replaces. remember(..., replaces=<id>)
  retires the old memory (kept, never deleted) and recall shows the new one.
  Undoing it through correction review brings the old memory back exactly.
  It works with a per-request profile_id too: a client routed to its own
  profile can replace a memory and undo it there (review_correction and
  list_corrections take the same profile_id), without switching the
  daemon's active profile.
• Upgrades keep a verified copy you can go back to: restore points, a
  restore that also brings back memories saved since the copy, and a prepared
  path back to the previous version (slm db restore-points | restore | prepare-downgrade).
• Encrypted cloud backups (see Security), with slm backup recovery-key and
  slm backup decrypt.

Fixed

• Recall never writes. Asking about an unknown name no longer creates
  entities or calls a model, which also removes a multi-second delay on the
  first question about a new name.
• Enrichment no longer gets stuck on memories that hold credentials.
• Memories blanked by an earlier release are repaired once, after a
  verified safety copy: their text, the facts searched from it, and wrongly
  set "replaced" marks.
• A restore never repeats a loss. It deletes again only what you or a
  privacy erasure deleted, works when the live store is damaged, refuses a copy
  made by a newer version, keeps kinds you confirmed after the copy, keeps
  erased profiles erased in learning data, and cannot be cancelled half-way
  through. Its messages say what actually happened.
• A kind you confirmed is never overwritten by background work, and a
  machine suggestion never changes a memory's type.
• The same request gives the same answer on MCP, the CLI and HTTP,
  including refusals. Every --json command prints JSON on errors too.
• The Hermes plugin works with the release it ships with, and asks for
  confirmation before remember --replaces.
• Windows: the test suite runs on Windows again, and a read timeout in the
  embedding worker no longer freezes it.
• npm install on Linux finds python3.12 and newer next to an older
  python3, and honours SLM_PYTHON.

Security

• Credentials leave this machine redacted on every path. One outbound gate
  covers model providers, cloud embedders, summaries, the reranker, the
  answer check, a LAN Ollama and mesh peers; a model provider's key is sent only
  to that provider. Context given to agent sessions is redacted the same way,
  with no last-four-character tails. Requests to this machine never go through
  an environment proxy.
• Cloud backups are encrypted on this machine before upload to GitHub or
  Google Drive. You get a recovery key; keep it to restore on another
  machine. Backups made by earlier releases still restore.
• Every new read route checks permission.
• Error responses no longer include file paths or internal messages.

Changed

• Credentials you save are stored as written. Up to 4.1.18, known key
  formats were removed on save. SLM keeps them so your agents can recall them,
  and redacts them on the way out.
• Python 3.12 or newer is required, and every check now says so.
• Tool sets: default and full 54 tools, code 38, power 66. The Antigravity
  plugin uses the power set (within Antigravity's limit) and the Cursor
  configuration uses the code set.
• Recall's ceiling is 3.0 seconds including the answer check.

Known limits

• Macs shared by several user accounts: the per-account check is still not
  in this release. Turn on company mode (Team → Access policy) on a shared Mac.
• Classifying existing memories with a cloud model uses on-device rules.
• Drive does not rotate old backups: delete plain-text backups made before
  this release yourself.