Sno Station

data_flow_transparency scores 3: security.md enumerates local SQLCipher encryption, key in settings.json at 0600, REM Enhanced and Agent Native default nightly session uploads, opt-out switches, and telemetry excluding memory content. least_privilege scores only 1: install is a curl|sh remote script and setup rewrites hooks, config.toml and permissions.allow across multiple agents with no minimization rationale. user_confirmation scores 2: Self Learning proposals require explicit accept and /memory clear requires --yes, but Agent Native's default nightly upload is on by default and must be disabled in settings.json, a weak confirmation posture. sensitive_data_handling scores 2: encryption and threat model are well documented, but key and store co-reside with no escrow, and REM Enhanced uploads plaintext memory text; boundaries are disclosed but real. dependency_security scores 1: many dependencies with caret ranges, no lockfile audit, SBOM or vulnerability scanning evidence; CI only builds. external_effects scores 1: install mutates several agent configs and registers hooks, a real side effect on the host with thin rollback guidance. rollback scores 1: docs stress reinstall preserves the store and to back up settings.json, but no version rollback or config-revert procedure. source_attribution scores 2: README credits Karpathy, WikiSkill and others, but code-level third-party attribution and NOTICE are not shown.

self_consistency scores 2: the three usage guides agree on mode tables, defaults and hook timeouts, and README matches security.md on REM; however README's 'no daemon' and 'runs on your laptop' sit in tension with an on-demand memory service and nightly uploads. dependency_availability scores 1: depends on Hugging Face model downloads (with mirror and SHA-256 pinning), npm packages, the Sno cloud and account allowance, limiting offline use. failure_messages scores 2: troubleshooting covers doctor states, no-repository-root, correction-in-progress and worker.log with concrete steps.

audience_and_scenarios scores 2: clearly targets developers running multiple agents, with quota handover, peer review and nightly learning scenarios. capability_boundaries scores 2: the Cursor section and supported-surfaces list name unsupported cases (Windows, Cursor cloud agents, VS Code extension). trigger_precision scores 2: hooks specify timeouts, character caps, short-prompt skipping and subagent exclusion. environment_fit scores 1: requires a narrow Node version range, a git repository, and is Linux/macOS-centric with macOS and Desktop marked untested.

information_architecture scores 2: README is well sectioned and docs are organized per harness, though numerous and densely cross-referenced. install_notes scores 2: commands, requirements, SNO_PROFILE_DIR and per-harness steps are present, but the core path is still a single curl|sh line without checksum or signature guidance. naming_stability scores 1: package names @snoai/mem-claude, mem-codex, mem-claw, mem-cursor coexist with sno/sno-mem commands, and the OpenClaw plugin name sno-mem-claw differs from its package name. examples_and_faq scores 2: example commands, an FAQ and described real-world evidence screenshots are provided. known_limitations scores 2: the 'what we do not protect against' list and Cursor's 'Not yet' section are candid. license scores 3: full Apache-2.0 text present and consistent with package.json. versioning_changelog scores 1: only 'Last updated' dates and a release badge; no CHANGELOG file. maintenance_responsibility scores 1: publisher identity is unverified and SECURITY.md only supports the main branch, leaving the responsible party unclear.

output_usability scores 2: commands and hooks produce concrete outputs (ids, supersede chains, JSON Lines export) usable directly. marginal_value scores 1: the core value (shared memory, peer review, nightly learning) depends on the Sno cloud and account allowance and overlaps with each agent's built-in memory, with independent gain not evidenced. cost_benefit scores 1: requires a resident service, rewrites of multiple agent configs and default nightly uploads, a non-trivial operational and privacy cost against largely vendor-asserted benefits.

claim_traceability scores 1: README claims 'proven end to end on a clean machine' but the repository shows only screenshots and evidence-directory references, not reproducible test scripts or logs. cross_source_corroboration scores 1: README, security.md and usage guides corroborate the data flows, but all are the same publisher's documents with no independent corroboration. fact_inference_separation scores 1: docs blend design intent with observed results (e.g. 'never once said looks good') without clearly separating fact from inference.