SIA — the Omarchy Brain

“Brain” is a product metaphor for auditable local machine memory; it is not a biological brain and does not establish cognition or neuroscience.

This machine has persistent local memory with semantic indexing. A daemon (sia-brainstem) tails the enabled evidence streams available here (base pacman/journal/git/session/notification sources plus optional JACKAL, SEKHMET/Custos/AEGIS, WORLDLINE, Guardian, and configured custom senses) into a git-versioned corpus indexed by gbrain with local embeddings. This session is a client of that memory.

CLI commands (a core subset is exposed by MCP server sia where mounted)

sia ask "question"        # semantic recall, cited, origin-labeled
sia ask "question" --no-touch  # audit/eval read that does not reinforce memory
sia recall <slug>         # one memory page verbatim
sia status                # readiness, chains, source summaries
sia ready                 # live exit-status gate for reconciled memory
sia think                 # recent generated entries
sia note "text" --from <you>      # leave a labeled note for future sessions
sia takes / sia calibration       # predictions and the Brier scorecard
sia take "claim" --confidence 0.8 --by DATE --domain d   # user-held take
sia take --accept <proposal-id>|all # commit immutable proposal(s) (HUMAN decision)
sia memory / sia memory <slug>    # stability, pin, and review state
sia rehearse [slug]               # list due reviews or deliberately recall one
sia bench                         # signed-ledger held-out memory self-benchmark

When to use

• Session start / before re-deriving machine history: sia ask first. Package changes, crashes, healings, chain verifications, prior agent sessions are recallable when their source was enabled, observed, and retained.
• When you learn something durable about this machine: sia note "..." --from <your-harness-name>. Notes are model-origin memories — recallable and graphed, permanently labeled agent prose, weighted below evidence. Write them for your successors, not as scratch. Never put credentials, secrets, or private content in a note: notes persist and can be returned to configured consumers, while SIA's redaction is pattern-based defense in depth rather than a secrecy guarantee.
• When you make a testable claim about this machine's future: propose a take (MCP sia.propose_take, or tell the user to run sia take). Agents propose; only the operator commits. the configured judge grades takes against recalled evidence when due; Brier scores use deterministic decimal arithmetic and remain descriptive over an operator-selected population.
• For an audit or evaluation: use the MCP sia.search tool or CLI --no-touch forms. Ordinary recall intentionally reinforces memory; the no-touch lane does not mutate rehearsal or graph state.

Honesty rules (non-negotiable)

• New persisted origin metadata has exactly three classes: [evidence], [derived], and [model]. Missing, malformed, or ambiguous legacy metadata is surfaced as [legacy-unlabeled]; it is never evidence and is weighted conservatively like [model]. Never cite model or legacy-unlabeled memory as evidence. Judge-grade/ponder entries, take-proposal notifications, and agent/operator notes are model; deterministic transition and Brier arithmetic do not relabel the verdict.
• Typed domain relations come from event/epoch evidence or an explicitly [derived] integrity/healing/crash/refusal generated entry. Model and legacy-unlabeled entries may link pages only with generic mentions.
• Every sia ask answer ends with a truth-boundary line (sense freshness, chain verdicts, recall mode). Repeat degradations to the user; never present an answer over stale SIA state as current.
• sia status remains available during recovery and labels its live readiness reason; its pulse and graph fields are the last published snapshot. Use sia ready when automation needs a success/failure gate.
• Absence of recall is not evidence of absence.
• Never touch ~/.local/share/sia/.gbrain/ directly. PGLite admits one owner, so every SIA-managed daemon/CLI/MCP/benchmark operation uses the same cross-process lease. Agent notes go through immutable request files and the brainstem acknowledges them only after corpus commit and index sync. This coordination is advisory, not a hostile same-user sandbox.
• SIA's MCP server also exposes byte-bounded sia://status, sia://thoughts, sia://calibration, sia://cortex, and sia://memory/{slug} resources; thoughts and cortex are compatibility URI names for generated entries and the root page.
• The local MCP server makes no cloud or external network calls; configured retrieval can use a loopback Ollama service. Its client receives the requested memory and may send it to a model/provider. The same trust boundary applies to scripts, pipes, or agents that capture sia CLI output.
• Open the cockpit from its bar widget; SUPER+SHIFT+B is available only when the operator explicitly enabled that optional binding. Full docs are at ~/.config/omarchy/plugins/khephri.sia/docs/MANUAL.md for an Omarchy install. For a standalone install, use docs/MANUAL.md in the checkout used to install SIA; if neither path remains, fall back to sia --help and the MCP tool/resource descriptions.