zer0dex Local Memory

The design uses local Ollama, Chroma, and project-scoped storage, requires the host to invoke retrieval explicitly, verifies a launch token and process identity before stopping a background server, and gives the publishing job only repository-read and PyPI OIDC permissions. Data flow, file locations, external effects, dependencies, authorship, and Hermes Labs attribution are clearly described. Deductions apply because the evidence shows no server authentication, encryption, secret scanning, or dependency-vulnerability audit, and the project expressly disclaims access-control and privacy-governance functions. Runtime dependencies have minimum versions but no lockfile or upper bounds. Server shutdown and stale-state cleanup provide partial recovery, but memory deletion, backup, and rollback of stored data are not demonstrated.

The README, package metadata, CLI surface, and tests describe a coherent product. Tests check version consistency and cover configuration round trips, argument validation, readiness, occupied ports, stale state, process identity, and local HTTP exchanges. Failure paths generally exit nonzero with actionable messages, including unreachable Ollama, a missing client, and zero extracted memories. Dependency availability is reduced because operation still requires a separately installed Ollama service and two downloaded models, constraints are broad minimums, and the supplied evidence omits the complete implementation needed to inspect every compatibility path.

The intended developers, scenarios, two-layer architecture, host responsibilities, and non-goals are explicit. Retrieval is precisely positioned as an explicit host-controlled lookup before model calls, and the adapter guidance keeps retrieved text in an untrusted context rather than installing an implicit hook. Collection, port, storage path, user, and model settings are configurable. Environment-fit points are deducted because CI evidence covers only Python 3.11/3.12 on Ubuntu; later Python versions are expressly untested, while the broader operating-system claim lacks equivalent support in the supplied files.

The README has strong information architecture covering preview, audience, rationale, installation, first success, integration, evidence limits, non-goals, development, citation, and licensing. Examples are substantial, and tests align naming and versions across metadata, runtime, citation, and machine-readable documents. The complete Apache-2.0 license, Alpha status, security contact, supported-version rule, and response targets are present. Deductions apply because there is no dedicated FAQ; the changelog and compatibility policy are referenced but absent from the supplied evidence; and although an author, organization, and security email are named, publisher identity is unverified and no broader maintainer roster or succession path is shown.

The query result is concise and usable as agent context, with memory text, score, and source. The dual-layer design offers plausible, clearly explained value over either a large flat file or opaque vector-only retrieval. The README reports recall on one 86-memory, 97-case workload and carefully limits generalization. Deductions apply because the underlying evaluation document is not supplied, so marginal value is supported only by the README summary, while model, memory, disk, installation-size, latency, and operating costs are not quantified sufficiently for a thorough cost-benefit judgment.

Many operational claims trace to package metadata, CI configuration, and focused tests; versions, commands, process safeguards, error behavior, and local exchanges are corroborated across files. The sample output is labeled illustrative, and benchmark observations are explicitly separated from general guarantees, earning strong fact-inference separation. Claim traceability is reduced because the full implementation, evaluation report, compatibility policy, and changelog are absent, preventing direct corroboration of performance figures and some documentation promises.