Obsidian Mind

The sources describe local models, index isolation, exposed roots, private-tag exclusion, read logging, repository-scoped memories, migration approval, and Git-based update paths in substantial detail. Tests also show rejection of traversal-shaped configuration, while prompts explicitly favor copying over deletion. Author, design influences, and MIT copyright attribution are clear, supporting full source-attribution credit. Deductions apply because ordinary workflows automatically write and reorganize extensive personal and workplace notes; cross-repository sessions can still read the vault directly, and the README concedes that preventing leakage into public PRs relies mainly on an instruction contract rather than enforced access control. The supplied implementation does not fully establish authorization boundaries for Slack, GitHub, or MCP access. Dependency evidence gives versions and fallbacks but no lockfile, audit policy, SBOM, hash pinning, or vulnerability-response evidence.

The documented architecture, commands, lifecycle, and fallback story are broadly coherent, while supplied tests materially support hygiene scanning, path filtering, missing-directory handling, deadline behavior, and reporting of cleanup failures. Lexical fallback without QMD, CPU fallback, diagnostics, and explicit Node requirements improve dependency availability. Deductions apply because many MCP, migration, classification, command, and cross-client behaviors are assertions without corresponding supplied implementation or tests. Failure-message evidence is concentrated in test helpers and hygiene hints and does not establish consistent actionable errors across installation, external integrations, and all write paths.

The material thoroughly addresses personal memory, project work, meetings, incidents, reviews, and cross-repository knowledge, with differentiated support statements for Claude Code, Codex CLI, Gemini CLI, and other clients. Boundaries are unusually explicit: QMD is optional, automatic memory loading is Claude-specific, folder-name identity can collide, affirmative MCP instructions may be ignored, model downloads are substantial, and the Node flag may change. Hook timing, classifications, exclusions, budgets, and degradation rules are precise. Environment-fit points are deducted because complete compatibility across the three primary clients and MCP is not corroborated by the supplied configurations or integration tests, and other clients are expressly only partially supported.

Information architecture, directory responsibilities, template fields, command naming, and daily workflows are documented comprehensively. Clone, ShardMind, QMD, upgrade, migration, dry-run, and approval procedures are concrete. Model sizes, performance costs, repository-name collisions, the experimental Node flag, and client-support differences are candidly documented, and the full MIT text is present. Deductions apply because there is no formal FAQ, some integration examples depend on omitted implementation, and CHANGELOG/version information is referenced but not supplied. Maintenance has a named author, contribution rules, PR-title enforcement, and update paths, but no clear support commitment, release ownership, or broader governance. Unverified publisher identity is treated as unknown rather than adverse.

The intended outputs are directly usable: structured notes, decision records, incident timelines, review briefs, indexes, backlinks, and hygiene warnings all have clear destinations, and tests support actionable wording for part of the hygiene system. Hierarchical loading, injection budgets, local models, keyword fallback, and tiered search costs make the cost-benefit story unusually transparent. Marginal-value credit is reduced because the broad benefits of persistent memory and automatic organization are principally product claims in the README; the supplied evidence lacks representative generated artifacts, comparisons, or static implementation coverage of the core end-to-end workflows needed to establish the claimed advantage over an ordinary Obsidian template plus agent instructions.

Some claims trace directly to the license, CI workflows, and named tests. Test comments also connect behavior to concrete failure modes and distinguish successful cleanup, recorded residue, rejected paths, and hygiene states that can return to zero. Deductions apply because most product capabilities are described by a single translated README, while the key MCP, privacy, migration, cross-client, and cost-recording implementations are absent from the supplied evidence, limiting corroboration. The product explicitly labels knowledge as verified, inferred, or unverified and discloses several measurement conditions and failure boundaries, but claims such as the store becoming more reliable as it grows, full cross-client support, and observed behavioral effectiveness are not consistently separated from inference or backed by independent source evidence.