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.

Sia: the Egyptian personification of perception, who rode the solar barque beside Hu and Heka.

Give your machine local memory. SIA stores a persistent local corpus, indexes it semantically, and compacts older day pages into weekly summaries. A resident daemon tails the evidence your machine already produces — package installs, journal errors, git commits, optional Git-backed Obsidian vault records, agent sessions, notifications, and any log you point it at — into a git-versioned markdown corpus, indexed into a typed knowledge graph with local embeddings. It retains, links, and retrieves admitted memories, and runs a deterministic nightly maintenance and compaction cycle. It helps you propose and commit falsifiable predictions, then records how they are graded. You can watch its generated-entry stream and query the memory it has admitted.

You can audit every stored word because the corpus is markdown in git and the daemon signs its own acts.

Current release: v1.8.0. Two receipts bind storage by device and inode number — the installer's corpus receipt and the delivery epoch's adoption — and a btrfs subvolume change, an rsync of the home directory, or a restore by copy changes those numbers. After that every pulse refused, sia ready reported a pending migration, continuity reported only that it "failed without reporting repository credentials", and the installer refused with "existing corpus receipt is invalid". Issue #12 reported it on btrfs; the maintainer machine hit it after a VM migration. Nothing is guessed now: every refusal names both identities and the remedy, sia readmit --yes re-binds the unchanged adoption through a sealed receipt that retains the previous identity, and SIA_READMIT_MOVED_STORAGE=1 ./install.sh does the same under the installer's own journal. First light on a CPU-only host is bounded by a finite, named, progress-reporting ceiling instead of an unattributed 120-second gate (issue #10); the pulse never runs back to back (issue #13); a moved log with identical bytes continues instead of replaying; and bounded helpers fall back, announced, to process-group isolation where a kernel refuses unprivileged PID namespaces — the reason the suite was red on Ubuntu runners while green on Arch. The last pulse failure is retained and shown by sia status, not only in journalctl.

Automatic Continuity never wakes the computer solely to run a backup. Its persistent timers catch up after the user session returns, serialize their work, and coalesce duplicate scheduled requests. The hourly job creates an encrypted recovery snapshot; the weekly job performs the deeper repository check and exact off-path restore verification. Inspect the live schedule at any time with sia backup schedule.

Example SIA cockpit. This static preview is not current machine state, a live readiness result, or evidence of cognitive performance.

Install (Omarchy)

SIA is already listed in the Omarchy Plugin Marketplace. Add its public plugin checkout with Omarchy's standard command:

omarchy plugin add https://github.com/AnubisQuantumCipher/sia.git --enable

The enabled surface detects that the resident memory service is absent and presents a SETUP gate. Open the cockpit from the SIA bar item and choose Begin first light. That explicit action asks this desktop to open a terminal and run SIA's fail-closed install.sh in it — loading or enabling the QML never runs the installer by itself. SIA holds that terminal open when the installer finishes, on success and on a named refusal; the cockpit reports whether the installer shell was actually observed to start and, once it has, steps aside so that terminal is in front. First light is substantial: it downloads pinned local toolchains and the embedding runtime, builds gbrain, pulls the pinned local embedding model, creates this machine's signing identity and empty corpus, and installs user services. Review the disclosure in the gate before continuing.

Installation is complete only when the installed runtime version matches the plugin, the first-light pulse publishes, and sia ready passes; until then the cockpit stays gated. After omarchy plugin update khephri.sia, an older resident runtime produces an UPDATE gate whose Finish update action runs the same visible installer. An interrupted or ambiguous generation is labeled REPAIR instead of being guessed; a resident record newer than the checkout shows AHEAD and disables installation.

Community plugins run unsandboxed as your user, and the git-URL form clones the current upstream branch rather than the commit last verified by the Marketplace — inspect the checkout before pressing the setup button. See the official development guide.

Privacy at install and after: ingestion, indexing, retrieval, and embedding make no cloud calls. The judge is disabled by default; only if you explicitly configure a Claude model does that separate tool-free CLI path send the recalled context you ask it to judge. Configured Codex CLI grading refuses because its documented read-only sandbox still permits reads. An MCP consumer is an operator-configured trust boundary: it receives the memory it requests over stdio and may forward that content to its own model/provider — the same boundary applies to any script or agent that captures sia CLI output.

The whole system, in one view

SIA is an owner-local, evidence-grounded memory architecture — not a cloud assistant and not an opaque database. The corpus pages, retained non-page migration provenance, and their Git history are the source of truth; gbrain/PGLite, local embeddings, and the cockpit graph are rebuildable projections. The resident brainstem alone materializes agent notes, and it publishes a generation only after the corpus is committed, index sync succeeds, and graph export succeeds; otherwise memory-dependent reads refuse with named projection debt instead of presenting stale memory as current.

machine evidence + explicit owner / agent notes
          │  bounded senses · redaction · origin labels · immutable note queue
          ▼
git-versioned corpus + retained provenance  ← source of truth / recoverable history
          │  one SIA-managed PGLite lease · publication barrier · live readiness gate
          ▼
local gbrain + PGLite + Ollama embeddings  ← rebuildable index + typed graph
          │
          ├── Quickshell cockpit + bar widget
          ├── local `sia` CLI
          └── stdio MCP tools and resources for resident agents

Ed25519 hash-chained lifecycle ledger: signed transitions and results

System layer	What SIA does now	What it does not silently claim
Evidence and integrity	Ingests bounded machine records and explicit notes; redacts secret-shaped spans; keeps evidence, derived, model, and legacy-unlabeled origins distinct; signs its lifecycle ledger; and gates memory-dependent reads behind sia ready.	Recall absence is never evidence of absence. A published graph snapshot is not a live readiness verdict.
Local retrieval policy	Builds local semantic recall and a typed graph; applies origin-weighted hybrid-query ranking, usage salience, co-return reinforcement, an optional graph reranker, novelty and intake-surprise scoring, a bounded attention window, weekly epoch compaction, stability decay, pins, and SM-2 scheduling.	These are deterministic policies and lexical link inferences — not evidence of human-like mental processes or real-world relationships.
Dated intents	Keeps operator-created sia intend commitments, surfaces them as deadlines approach, and closes them only on the operator's word; a nightly slug-retrieval drift tripwire signals when to run the full benchmark.	It does not infer task completion, and its drift signal is a heuristic — not an answer-quality score.
Prediction calibration	Lets people commit future-dated predictions; optionally obtains tool-isolated evidence judgments; records signed grades and population-aware descriptive Brier calibration; and runs sia bench, a signed-ledger QA benchmark that scores abstention.	Calibration is not a representative population claim, and sia bench is a local regression instrument.
Resident agents	Exposes local memory through bounded stdio MCP tools/resources and sia context packs. Agents queue immutable note requests; the brainstem alone materializes and indexes them before acknowledgement.	Agents never receive a database handle. Their notes remain model-origin prose, and each MCP consumer is its own disclosure boundary.
Mission control	Presents the graph, generated-entry stream, evidence chain, source health, memory lens, agent relay, and an on-demand live-readiness check in the Quickshell cockpit.	The cockpit labels last-published diagnostics separately from an explicit live check, and labels bounded display omissions rather than implying memory is missing.
Continuity	Freezes the documented SIA roots into a signed portable capsule, verifies repository copies by exact off-path round trip, and thaws only through a journaled, receipt-preserving restore ceremony.	The private signing key and .gbrain are not in routine capsules. No repository is pruned automatically, and a same-disk copy is not disaster protection.

New here? Install · Try it · Field Manual · Whitepaper

What you get

• A corpus that accretes. Every admitted event becomes a durable record in a git-versioned day page; the corpus is authoritative and the database is a rebuildable index — wired into a typed knowledge graph by gbrain with local nomic-embed-text:v1.5 embeddings via Ollama. When a schema pack is unsafe or a projection fails, the affected edges degrade honestly and publication debt keeps memory-dependent reads closed until a pulse publishes a complete snapshot. The exact two-lane link-inference rules and their refusal behavior are specified in the Field Manual.
• A policy layer, not just an index. Deterministic policies adjust local retrieval: usage salience decays with time and grows with world-originated use; co-return edge weights increase when memories are returned together, with nightly decay and degree caps; graph propagation uses Personalized PageRank, measured on the tripwire probes and currently default-off in sia ask because the extended set showed it trailing the unmodified hybrid query — the whitepaper records the numbers and the re-promotion condition; a non-destructive stability lens demotes stale associations without deleting evidence, while safety-priority or operator-pinned pages follow a nightly SM-2 scheduling schedule; novelty and unusual sampled event intake produce generated notices without treating unobserved hours as source silence; a seven-entry attention window holds recently used memories; old day pages compact into weekly epoch summaries while declared safety-class days remain verbatim. Every compacted original remains recoverable in git.
• Prediction calibration. Register falsifiable predictions with confidence and strictly future UTC deadlines; a tool-free Claude judge (off by default) grades them strictly against recalled evidence without seeing your confidence — TRUE / FALSE / UNRESOLVABLE, abstention audited. Calibration reports are population-aware: a lone grade is labeled a single case, sparse bins are withheld, and the series stays descriptive because takes are not a random sample. Agents may propose predictions; only you commit them.
• Dated intents. sia intend "rotate the keys" --by 2026-10-01: commitments SIA surfaces as deadlines near and marks urgent when overdue, closing only on your word — a diary lane with no scoring and no model.
• A mission-control cockpit. A full-screen Quickshell overlay (from the bar widget, or SUPER+SHIFT+B when you opt into the binding): the animated graph with radial time, stable semantic sectors, collision-aware labels, replayable outward growth, hover neighborhoods, edge explanations, origin labels, a generated-entry stream, evidence-chain verdicts, and a SOURCE HEALTH truth boundary that admits incompleteness instead of hiding it. Its truth ribbon separates the last-published snapshot from an explicitly requested live sia ready check, and a reversible workspace lock keeps the cockpit on one focused Hyprland workspace. Plus a bar widget with the live event count.
• Agents everywhere. An MCP server for the documented Claude Code, Codex CLI, and Grok integrations (plus compatible stdio MCP clients), and a skill for skill-reading harnesses. Tools cover reinforcing recall, a read-only search lane for audits, and carefully labeled writes; resources mount status, generated entries, calibration, the root page, and sia://memory/{slug} pages (thoughts and cortex remain compatibility URIs). The clean client-visible tool surface is sia.ask, sia.search, sia.recall, sia.status, sia.think, sia.note, sia.propose_take, and sia.calibration; reconnect the client or open a new agent session after an update so cached MCP discovery is refreshed. Notes persist and may be returned to configured consumers, so do not put credentials, secrets, or private content in them; pattern-based redaction is defense in depth, not a secrecy guarantee.
• Evidence culture. SIA keeps its own Ed25519 hash-chained run ledger; every sia ask answer carries a truth-boundary line and one of the three canonical origin labels: evidence / derived / model. Ambiguous legacy origin metadata surfaces as the explicit legacy-unlabeled boundary, never promoted to evidence. Judge grades, ponder output, and agent/operator notes are model and never mint facts. Secret-shaped spans are redacted at the sense boundary; absence of recall is never evidence of absence.

The historian is the product; policy evaluation is the research program

The historian half — origin-labeled capture, the git corpus, local recall, the signed ledger, the cockpit — is shipped, tested behavior you can use today. The candidate policy layer — usage salience, co-return reinforcement, graph propagation, novelty scoring, a bounded attention window, stability, and rehearsal — is shipped as deterministic behavior, and the whitepaper is explicit about its evidentiary status: the associative tie-breaker matched the unmodified hybrid query on the historical probe set and did not beat it, and this release contains no controlled evidence that rehearsal improves answer quality. The policies are explicit hypotheses with instruments attached (sia bench, the nightly drift tripwire, the calibration record), not established improvements. Keep evidence and model separate in the docs the way the corpus keeps them separate in memory.

Sixty seconds after install

sia status                          # current source and memory-index state
sia ready                           # exit nonzero unless memory is reconciled
sia ask "what happened today"       # semantic recall, cited + labeled
sia think                           # recent generated entries
sia context                         # bounded handoff pack for agents/sessions
sia take "the build will go green" --confidence 0.8 --by 2026-09-05
sia intend "rotate ledger keys" --by 2026-10-01   # dated commitment
sia note "hard-won context" --from me    # a memory for future sessions
sia memory                          # stability, pins, and reviews due
sia memory --pin organs/journal     # protect/qualify a page for rehearsal
sia calibration                     # population-aware descriptive scorecard
sia bench generate --out /tmp/sia-qa  # signed-ledger QA + private MCP eval
sia bench generate --cognitive-history --live-history --out /tmp/sia-history
                                     # private signed + controller-use captures
sia backup status                   # continuity adapter and verified-copy state
sia backup schedule                 # authenticated automatic-timer state

Point it at your own programs in ~/.config/sia/config.json:

{ "custom_senses": [
    { "name": "myapp", "path": "~/logs/app.log", "type": "lines",
      "match": "ERROR|FATAL", "kind": "error", "tags": ["failed"] } ] }

match takes a bounded list of literal substrings separated by |; regular-expression operators are refused so a configured pattern cannot monopolize the resident writer. For type: "jsonl", SIA admits only the exact configured field. A real Git repository at ~/Obsidian is discovered as an optional source adapter with no configuration; for a different vault path, skill discovery roots, and judge settings, see the Field Manual — then restart the brainstem and check sia status.

Requirements

Linux (Omarchy/Arch tested; x86_64 or aarch64) with procfs mounted at /proc and kernel support for pollable pidfds, child-subreaper adoption, and sealed memfd files. python3 must expose os.pidfd_open, signal.pidfd_send_signal, os.waitid with waitid/WNOWAIT/WSTOPPED, and os.memfd_create, and must have an Ed25519-capable Python-cryptography. The native lifetime owner also requires Linux SCM_CREDENTIALS and SCM_RIGHTS; absence of any core lifetime feature refuses before the mutating shell launches. Bounded helper processes (chain keepers, gbrain) run inside a private user+PID namespace when the kernel permits unprivileged unshare; where it does not (Ubuntu 24.04's AppArmor default, hardened kernels) SIA announces once that it fell back to process-group isolation and keeps working — that fallback is weaker containment, not a failure. SIA also requires git, curl, tar, unzip, bzip2, sha256sum, zstd, flock, ss from iproute2, a systemd user session (systemctl), and roughly 2 GB of disk for Ollama. The bootstrap downloads Bun, Ollama, and SIA's private restic executable from pinned release URLs and verifies their published SHA-256 digests before extraction; it checks out the full gbrain commit in GBRAIN_PIN, verifies the pinned bun.lock, installs with --frozen-lockfile, compiles the executable, and receipts the result. Managed filesystems must support atomic rename and renameat2(RENAME_NOREPLACE); the share filesystem must support O_TMPFILE for crash-closed signed-ledger publication — the installer probes both before activating anything. Optional: the Omarchy 4.x shell for the cockpit. Judge calls remain off until you set judge.backend to claude with an explicit judge.model. ollama.service accepts exactly one operator-owned drop-in at ~/.config/systemd/user/ollama.service.d/sia-operator.conf — a [Service] header followed only by Environment=OLLAMA_<NAME>=<value> lines, never OLLAMA_HOST or OLLAMA_MODELS (SIA manages those) — for hardware knobs Ollama does not otherwise expose, such as OLLAMA_IGPU_ENABLE=1 when it would otherwise drop an integrated GPU (issue #10). SIA never writes this file or enables a hardware setting itself; any other drop-in, or a malformed one, refuses the install.

Install, recovery, and removal boundaries

The short version of the contract; the full mechanics are in the Field Manual and each release's CHANGELOG entry.

• Failures stay loud and recoverable. Before any dependency mutation, a durable launch fence binds the old CLI, brainstem, and MCP launchers and sets them to mode 000 so nothing enters through a stale launcher mid-upgrade. A failure after the first installer mutation deliberately leaves the brainstem disabled and stopped instead of restarting it against mixed dependencies — fix the reported cause and rerun install.sh. Prior runtime and plugin trees are retained at printed hidden sibling paths.
• Readiness is a live gate. Memory-dependent commands hold the corpus lease from the readiness check through the returned result, keeping the answer in the same corpus generation, and report SIA memory read refused until a successful sia pulse reconciles named publication debt. The sia status readiness line is live; pulse/graph fields are last-published snapshots. Never delete a sync_needed marker or a pending grade-transactions journal to "fix" readiness.
• Corpus creation and adoption are attributed transitions, not directory-shape guesses: an absent corpus is assembled off-path and advances through prepared/publishing/published under a durable managed-install/corpus-bootstrap record; adoption writes a corpus-adoption receipt bound to the directory's stable identity. An unreceipted corpus is recognized as legacy SIA only when .git/ is real, README.md contains the exact marker line # SIA corpus — this machine's memory, and bin/sia-ledger verify accepts the share tree; other nonempty corpora require explicit SIA_ADOPT_EXISTING_CORPUS=1 consent. The runtime receipt hashes only the allowlisted shipped members; it does not attest to extra entries. Preserve an interrupted bootstrap or adoption record and rerun the installer — deleting it does not grant adoption authority.
• The signed ledger initializes as an exact prefix: key.hex, then the matching pub.hex, then one canonical signed GENESIS:init row in ledger.tsv, then the matching head.pin. Any other combination refuses closed; do not delete or recreate individual ledger components.
• The gbrain store bootstraps through its own supported front door under a durable managed-install/gbrain-bootstrap transition. Because gbrain records an absolute database_path, the authenticated probing phase generation-CAS rebinds only the exact private bootstrap path to the canonical brain.pglite path before the final health probe (the fix for issue #2). Unhealthy or unattributed stores are preserved and refused; direct .gbrain rebuilds are unsupported.
• Continuity restores are ceremonies, not copies. Restore preserves the installed destination .gbrain root, config, schema pack, managed receipt, and unknown children, rebuilds only the PGLite projection through gbrain, and turns green only after a fresh correlated health, ledger, and adoption proof. SIA's private restic adapter uploads a verified capsule hourly and round-trip-verifies weekly; it never prunes, and routine capsules exclude the private signing key. See Continuity. (In this repository SIA means only the Omarchy Brain product — not the Sia Foundation's sia.tech storage network.)
• Plugin enablement refuses rather than guesses: the installer requires an exact khephri.sia catalog entry after a rescan and refuses malformed/non-list JSON; a stale Quickshell entry after uninstall is cleared with omarchy-shell shell rescanPlugins.

Documentation

• Field Manual — cockpit tour, full CLI, generated-entry glyphs, configuration, the exact operating and recovery paths, troubleshooting.
• Continuity — signed portable capsules, storage adapters, automatic verification, and clean-machine restore.
• Whitepaper — architecture, the evidence model, every candidate policy with its formula and literature ancestry, the measurement instruments, and the verification record.
• Roadmap — the measured state of every instrument, the feature freeze and its exit gates, the module-split schedule, and the v2.0 decision rule for the candidate-policy lane.
• CHANGELOG — the complete release history. Recent highlights: the v1.5.1 rehearsal-grading fix with its real-gbrain contract test lane (issue #3), v1.5.0 marketplace-native guided first light, the v1.4.2 fresh-bootstrap recovery reported by @m10ust, and the v1.4.1 optional Obsidian source adapter proposed by @webdevtodayjason.

Omarchy marketplace status

SIA is listed as khephri.sia. Omarchy has no install/update/remove lifecycle hooks, so SIA presents an explicit cockpit gate and, only after you press the setup/update button, asks the desktop to open a terminal running install.sh; the cockpit reports whether that installer shell actually started. The Marketplace page may continue to display Manual setup until a maintainer removes its registry override; that catalog label does not change the repository's guided flow. Every release is validated locally with omarchy plugin validate . and listing updates go through the marketplace's newer-upstream-commit verification path. Marketplace validation and listing are not security reviews.

Remove

Use the installed teardown command:

sia uninstall           # removes code/UI; keeps corpus, ledger, keys, queues, config
sia uninstall --purge   # also attempts to erase retained SIA data and config

The stable sia launcher and sealed uninstaller are runtime members, so this route remains available if Omarchy has already removed the plugin checkout. The resident brainstem observes the canonical plugin registration before each pulse and stops intentionally, before more memory work, when it disappears. Repository checkouts also retain ./uninstall.sh as the source-tree entry.

On success, either path disables the Quickshell surface and archives the plugin checkout, so a later omarchy plugin remove is normally unnecessary. A plain omarchy plugin remove khephri.sia used first removes only the checkout; it does not uninstall SIA's resident runtime or user service. Run sia uninstall afterward to finish the retained-data teardown.

Default removal preserves the data categories — corpus, ledger, signing identity/head, queues and state snapshots, research, private toolchain, and operator config — not byte-for-byte intact directory trees; it archives the active plugin tree to a printed hidden sibling path instead of deleting possible local additions. Ollama, its user service, and the local model remain because they may be shared. If the brainstem cannot be stopped, removal preserves its runtime and unit, and --purge is blocked rather than deleting memory under a possibly live daemon; surviving or indeterminate service/plugin/MCP consumers likewise block purge, while an unrelated mismatched MCP registration is preserved without retaining SIA's runtime. --purge is destructive but not transactional: it attempts the state, share, and config root removals independently with no rollback — back up every retained category first and inspect the aggregate result. Review uninstall.sh before using --purge if the corpus or signing keys have not been backed up.

What this is (and is not)

A local, git-versioned, origin-labeled memory that refuses to pretend a language model is a witness. It is a disciplined machine-event historian with a small semantic index and a cockpit. Product policy names describe deterministic behavior; literature ancestry is kept in Credits and References and is not a warrant for stronger claims.

Honesty principles (the actual design)

1. Every recall answer (sia ask/search) declares its origin and truth boundary.
2. Absence of recall is not evidence of absence.
3. A system that fails open must say so (SOURCE HEALTH).
4. The model may summarize and grade — never mint facts. Its grades, ponder output, and agent/operator notes remain model even where SIA separately applies deterministic transition or Brier arithmetic.
5. Records, not content: built-in senses use metadata and evidence streams; secrets are redacted at ingest, and built-in senses never open agent message bodies, clipboards, or private keys. Operator-configured custom senses read the file/field explicitly named in config and must not target secret or content stores. The separate ledger keeper necessarily reads SIA's own signing key when it signs an authorized transition.
6. SIA does not silently delete or guess: weekly compaction is git-recoverable, named lifecycle transitions (boot, pulse ingests, maintenance runs, and grades) are ledgered, and refusal/partial states remain explicit. Graph read/export failures are visible in SOURCE HEALTH; a pulse that cannot publish its graph records the signed result graph-fail.

Credits

Built on gbrain by Garry Tan. The fresh-machine bootstrap recovery was reported by @m10ust in issue #2; the optional Git-backed Obsidian source adapter was proposed by @webdevtodayjason in issue #1. Candidate-policy ancestry includes Anderson (ACT-R), Collins & Loftus, Nader, Lisman & Grace, McGaugh, McClelland/McNaughton/O'Reilly, Dehaene, and to HippoRAG, Generative Agents, Zep, and Letta — citations in the whitepaper; the names are ancestry, the behavior is the contract.

MIT © 2026 Khephri Labs