AgentMemory for Codex on Windows is an independent, Windows-native AgentMemory downstream project, targeting OpenAI Codex Desktop and Codex CLI. It is based on AgentMemory v0.9.30, but is not the official upstream repository, an @agentmemory/* npm release, nor does it represent an upstream support commitment. The project is positioned as a technical preview, generated by AI and tested by users, with relevant disclosure documents available for review.
In terms of purpose, the project aims to retain useful context for Codex tasks while treating the official AgentMemory memory, lessons, graph, audit, and provenance stores as the sole authoritative storage. It captures normal main-agent user prompts and final assistant replies through four managed Codex hooks (SessionStart, UserPromptSubmit, Stop, SessionEnd), and excludes ambient UI, title/fork traffic, known internal host prompts, and subagent traffic, avoiding persistent capture of irrelevant content. Writes, deletions, curation, and provenance are strictly limited to the exact project scope.
Main capabilities include: support for audited, target-scoped live graph provenance corrections, with dry-run previews; support for audited recovery of observation records proven to be empty via the existing REST forget endpoint, preserving exact IDs, version checks, and graph cursors; support for bounded, source-labeled federated recall across projects, but wildcard writes are not allowed; typed graph extraction using a credential-free, loopback-only local Qwen worker, with other AgentMemory LLM features receiving a noop provider and external fallback remaining disabled; use of an authenticated loopback MCP endpoint for Codex, while retaining the packaged stdio launcher as a compatibility path. The upstream-compatible source surface includes 58 MCP tools, 6 resources, 3 prompts, 143 endpoints on port 3111, 12 portable hooks, and 17 skills, but the supported Windows profile intentionally activates only the four managed hooks mentioned above.
In terms of usage, users do not need to install upstream first. Before installation, check Windows x64, Node.js 24+, Windows PowerShell 5.1, and Codex Desktop, then follow the fixed installation guide. Released versions include prebuiltAt runtime, building from source is optional. The installer performs a dry run by default, validating the release hash, ownership, paths, and existing installation, and only executes with --execute after everything checks out. For an empty target, the default path prepares, activates, and verifies the service; --fresh and --activate-prepared can be used to separate the stages. After service verification passes, connect the Codex MCP. If a step fails, refer to the support and recovery guide, report the release version and failing stage, and include a brief redacted error, but do not attach the memory database or credentials.
The applicable scenarios are mainly developers who need project-level persistent memory and bounded cross-project recall for Codex on Windows. Regarding search and filtering, normal recall searches stored observations and reusable lessons; keyword retrieval uses BM25 and handles Korean morphological variation and literal identifiers; when available, vector and graph store candidates are added to the same ranking result set. Graph search reads existing nodes, edges, and source references and does not require the local Qwen to answer queries; Qwen is used only for background graph extraction. Search candidates are checked against their authoritative observations and sessions, and project, speaker, archive, and environment source eligibility is applied before determining the number of results returned. Literal code and quoted source text are preserved.
Regarding limitations, this preview is an unsigned technical preview, the binaries are not Authenticode-signed, and SHA-256 verification only verifies artifact integrity, not publisher signature. AdGuard/redirect driver interoperability has not been qualified. Validation on one Windows/Codex host does not represent all host configurations or all upstream-compatible operations. There is a known intermittent caller response delay of about six seconds: native MCP completion takes about 1.4 seconds, after which it takes about 4.5 seconds for the calling script to resume; the exact cause is unknown. The graph ranking fix is another confirmed defect and is not a fix for that delay. Validation scope covers package behavior, managed hooks, capture/filtering, authenticated MCP/REST, graph provenance, installer/recovery, and the packaged CLI/MCP entry points, but exact release results must be checked in the release notes.
