Mnemoverse Memory

The README clearly discloses the hosted-by-default service, authentication paths, shared-room boundaries, and the conditions for self-hosting. Key-entry guidance avoids shell-history exposure, while the supplied tests verify that unsafe non-TLS remote URLs are rejected before transmission and credential-bearing URLs are not echoed. The tool surface is constrained and Vault values are explicitly unavailable, but writes, feedback, room creation, and invitations still change external state without evidence of a uniform per-operation confirmation mechanism. Dependencies are few, yet no lockfile, vulnerability scan, SBOM, or dependency-update policy is supplied. User-facing deletion has been removed and no undo, recovery, or export path is documented, so rollback scores zero. Repository, author email, research, and licensing attribution are present, but the publisher identity remains unverified and no stronger identity inference is made.

The README, package metadata, and tests are strongly consistent about the ten tools, room isolation, feedback semantics, and error boundaries. Assembled-response tests cover empty results, failed probes, malformed responses, and name fidelity; URL-guard tests verify refusal before any request, supporting full scores for consistency and failure messaging. Dependency availability is reduced to 2 because normal operation depends on an external hosted engine, network access, authentication, and npm; the README acknowledges stuck connector sessions requiring reconnection. Daily contract-freshness checks detect drift but do not remove those dependencies.

The material gives tailored configurations for Claude Code, Cursor, VS Code, Windsurf, Zed, JetBrains, Cline, Continue, and remote OAuth, while addressing personal continuity, shared rooms, and self-hosting needs. Tool descriptions and tests clearly delimit personal domains versus rooms, read-only membership, pagination, unusable filters, and importance-gated writes. Trigger precision loses one point because this repository mainly exposes tools and prompts; the broader discipline governing when an agent should proactively recall or save is delegated to a separate project rather than established here.

The README has strong organization and concrete installation, key validation, client configuration, smoke-test, tool, prompt, resource, use-case, and troubleshooting material. Compatibility rules describe patch/minor stability, and tests constrain naming and advertised descriptions. The complete MIT text agrees with package metadata. Versioning receives 2 because PR and release-gate discipline is described, but the actual CHANGELOG contents are not included for inspection. Repository ownership, contact details, funding, and update automation are visible, though no support commitment, response policy, or independently verified responsible entity is established.

Outputs are designed for direct model and user consumption: errors are attributable, search scope is disclosed, names remain reproducible, and empty results avoid overclaiming; detailed static tests support these properties. Cross-tool persistence, shared rooms, and feedback-driven reranking offer plausible value beyond per-tool files or similarity-only retrieval, but the evidence supplies no comparative experiment or benchmark, limiting marginal value to 2. Hosted operation and a stated free tier may reduce operational cost, yet quotas, prices, latency, retention, and migration costs are absent, so cost-benefit is only adequately documented.

Many client-side claims trace to README text, package metadata, workflows, or explicit test assertions. The material carefully distinguishes the open-source MCP server from the closed hosted engine, enabled from disabled features, and current facts from future benchmark claims, justifying full fact/inference separation. Deductions reflect that the hosted engine implementation, live connector parity, registry listing, research conclusions, and operational behavior cannot be independently verified from the supplied files. Cross-file corroboration exists, but the sources remain predominantly publisher-controlled rather than independent.