mex-memory/mex

npx mex-agent@0.8.4 setup

One engineer knows why a constraint exists. Another has the debugging history. A coding agent found an important edge case in a session nobody else will read. The next teammate has to piece it together again.

What one engineer and their agent learn should become context the next teammate can use. MEX gives that knowledge a durable home in the repository: readable Markdown, code-linked explanations, reviewed knowledge contributions, and structured handoffs. People explore and review it in a local Hub; agents retrieve and help maintain it through project instructions and the CLI.

💬 Join the MEX community on Discord — discuss ideas, get help, share feedback, and show what you're building.

Join the Discord →

Canonical memory travels with ordinary Git commit, push, and pull. Each teammate keeps their own local indexes, drafts, identity selection, and Hub. No hosted MEX service, Docker, proxy, MEX account, or MEX-owned model key is required.

Working solo? The next person using that memory can be you in a new session.

An example: Alex changes webhook retry handling, and Sam will continue the work. Both are active MEX Members in a repository their team has already set up.

1. Start with the team's context. Alex asks Codex to inspect the existing architecture, relevant decisions, and code evidence before making the change and running tests.
2. Keep the useful discoveries. With Alex's direction, Codex updates the relevant Wiki explanation and code references. For a conclusion the team should review, it prepares an Inbox knowledge proposal for explicit approval.
3. Prepare and publish the handoff. Alex asks $mex-relay to draft a Relay for Sam: what changed, which tests ran, what remains, and where to look next. She reviews the draft and publication preview in Hub, explicitly publishes it, then reviews, commits, and pushes the code and canonical MEX files through Git.
4. Continue from shared context. Sam pulls the relevant branch, updates his local indexes as needed, and opens Hub. He reviews and takes the Relay, then asks his coding agent to read its context and continue. His acknowledgement is another canonical change to commit and push.

The Relay carries the explanation and observed repository state—not the uncommitted code. Publishing writes files to Alex's checkout; it does not notify Sam or deliver anything until they share through Git. See Relay boundaries for lifecycle and concurrency details.

The Hub is where people explore and review their team's memory. Open it to explore connected project knowledge and code, review a proposed knowledge change, find an eligible handoff, or see recorded team history.

• Understand the project: Home, Search, Context, and Code bring explanations and implementation evidence together. The Context graph shows knowledge entities and relationships; selecting one reveals its direct code groundings and details.
• Review and carry work forward: Inbox proposes additions and corrections to project knowledge; Relays preserve what the next person needs. Existing Spec proposals and Workstream records remain readable.
• See who and what: Team/Members supports attribution and local identity selection. Activity shows accepted MEX workflow events and recorded project notes—not every code edit or Git action.
• Keep context usable: Health and Jobs expose index status and explicit maintenance.

After setup, run mex hub. Each engineer's Hub reads their own checkout and listens on 127.0.0.1; it is not a shared hosted dashboard. Git brings the team's canonical records into that checkout. Hub protects mutations with a server-side session and CSRF token. Playbook and Catch Up product workflows remain future work.

MEX requires Node.js 22.5 or newer, SQLite FTS5 support in that Node build, and a Git repository. See runtime compatibility. The normal npm flow works on macOS, Linux, Windows Command Prompt, PowerShell, and WSL.

Run from the repository root:

npx mex-agent@0.8.4 setup

This opens a keyboard-driven terminal HUD with a cyan ASCII mex banner. Highlight a tool with the arrow keys and press Enter to select it and begin. Space optionally selects additional integrations; the Continue row starts with your checked choices or without an agent. Long steps show an animated activity indicator, elapsed time, and the latest reported work; graph parsing also shows real file counts. When a selected Claude Code or Codex CLI is available, its normal interactive session takes over the terminal; exit that session to return to MEX. MEX checks the scaffold before continuing. Incomplete work can be reopened, populated manually, or left for later. Existing instructions are preserved.

After validation, setup opens the Hub for completion. Review the exact setup-file diff and choose Commit setup to create a local commit. A manual Git checkpoint remains available. MEX preserves unrelated staged work and never pushes or pulls.

The completion screen explains how to start a fresh agent session and verify that it reads your project memory. It also offers two optional steps: install the mex command globally at the version running setup, and leave an email (plus an optional name) for follow-up about MEX. Contact details are sent through the embedded Web3Forms service and stay out of the repository and usage telemetry. Only a submitted/skipped preference is remembered on this computer. Choose Open Hub when ready; the Hub's introductory tour starts there.

Connected agents have their own installation, account, and network requirements. Keep the setup terminal running while using the finishing Hub; press q or Ctrl+C to stop it. Use npx mex-agent@0.8.4 setup --no-open to show the finishing link without opening a browser; add --port <n> to choose a loopback port. Press d in the HUD to read longer status or error details.

Expected setup errors include specific recovery instructions, such as correcting malformed configuration or an ignore rule that hides the scaffold. An incomplete agent session returns to recovery choices; a successful agent exit alone does not mark setup complete.

Prefer browser setup?

npx mex-agent@0.8.4 setup --browser

The browser runs the same setup engine with a background Claude Code or Codex session and manual population fallback. --browser --no-open prints its link without launching a browser. The existing --cli flag remains an alias for the default terminal flow.

For scripts, use --yes with repeatable --tool <name> flags (or --tool none). Piped input/output and CI also use plain output. These runs reuse saved choices, default to no integrations when none are saved, and never launch an agent or browser or wait for input. Exit codes are 0 for completed preparation or a successful dry run, 2 when population is still required, 1 for failure, and 130/143 for interruption. After a population checkpoint, rerun mex setup to continue. Once preparation is complete, mex setup --browser opens the commit and completion page.

mex setup --dry-run remains a terminal preview without changes. Once installed, bare mex opens the Hub for an established code project or terminal setup for an incomplete project. mex hub explicitly opens the browser, while mex tui opens the terminal dashboard.

Push the reviewed setup commit through your team's normal Git workflow so teammates receive the same project memory and selected agent instructions. In Hub's Team/Members page, add the people who will participate and choose your local identity. Review and apply those actions explicitly; commit and push new Member records too. Your current-member selection stays local.

Clone or pull the team's repository and branch through Git. For a completed, committed 0.8 setup, build the derived indexes in your own checkout and open Hub:

npx mex-agent@0.8.4 graph rebuild
npx mex-agent@0.8.4 wiki rebuild-index
npx mex-agent@0.8.4 hub

Reuse the shared project memory; do not regenerate it just to join. In Team/Members, check the effective identity and, if needed, choose your existing Member record as a local override. If you do not have a record yet, explicitly create one through the reviewed workflow and share its canonical files. Members are attribution, not a sign-in or permission system.

The committed instruction files and skill directories are reusable by the agents they target. Install the agent separately and start a new session in the repository. If your chosen integration was not included in the shared setup, coordinate adding it with the team; see agent integrations. After later pulls or branch changes, inspect Graph/Wiki health and run the indicated explicit maintenance—reads do not silently update indexes.

Older or incomplete setups should follow upgrade and compatibility first. Keep CLI versions aligned before exchanging new Relays.

npm install -g mex-agent@0.8.4
mex setup

npx mex-agent@0.8.4 setup --mode agent-memory

Examples use mex for readability. Install it globally as above or replace it with npx mex-agent@0.8.4.

The team's memory is shared; the machinery that retrieves it stays local. MEX separates canonical repository files from rebuildable indexes so each engineer and agent can work against their own checkout.

Canonical knowledge is structured Markdown with metadata, relations, sources, provenance, and code groundings; accepted Wiki writes append audit records. The Code Graph and Wiki search index are rebuildable local SQLite views, not shared sources of truth.

In 0.8.1, ordinary Wiki creation and synthesis capture the operation's recorded actor, time, and session as creation provenance while preserving any supplied original attribution. Migration leaves unknown, untyped context/*.md files untouched; declare the file's type with a root type key in its frontmatter (for example type: component) before migrating it, instead of relying on an architecture default. mex check and mex wiki validate report knowledge files that are not yet in the Wiki.

Git carries the meaning; setup or explicit maintenance commands rebuild the indexes against each checkout's own branch and working tree. Code remains authoritative, and grounding drift flags code-linked explanations that need review.

Shared memory needs both the team's explanation and evidence from the implementation. MEX combines two complementary views of a repository:

• The Wiki explains architecture, conventions, decisions, patterns, topics, and Specs in language people can review.
• The Code Graph uses bundled Tree-sitter grammars to map symbols and relationships from the implementation into a local SQLite index for precise, bounded retrieval.

Inspect or explicitly maintain them:

mex graph status
mex graph refresh       # Republish an existing compatible store
mex graph rebuild       # Full replacement when status requires it
mex wiki rebuild-index
mex wiki query "authentication"

Ask the Graph for a task-sized evidence set or an exact structural relationship:

mex graph scope "trace the authentication flow"
mex graph query where-defined authenticate
mex graph query who-calls requireSession
mex graph get <node-id>
mex impact requireSession

MEX indexes TypeScript/TSX, JavaScript/JSX, Python, and Rust, with partial C# support. Module variants such as .mts, .cts, .mjs, and .cjs have partial coverage. Bounded framework resolvers cover Express, Next.js App Router, FastAPI, Flask, and NestJS; see the support matrix for their limits. CLI Graph reads can return bounded, clearly labelled degraded evidence when configuration drifts, parsing is incomplete, or changed files must be excluded. Incompatible engines still refuse reads, and Hub Code keeps strict freshness checks. scope can also return bounded live-text evidence for stale or unindexed files, marked text-only.

Hub Graph construction runs in a disposable process so compiler work can release its memory when the job exits. Refresh reuses unchanged extractions and rewrites changed rows, while still resolving the whole corpus; broad compiler or configuration changes fall back to full extraction. Process isolation improves responsiveness and releases compiler state after the job, but does not guarantee lower combined peak memory.

A Wiki claim can point to a deterministic graph node. MEX stores the node ID and identity fingerprint; new MEX-written groundings also carry a body hash, while compatible legacy groundings may fall back to coarser fingerprint comparison. Together these signals distinguish intact, changed, moved, missing, ambiguous, and unverified references.

Drift is a review signal. It does not prove that prose is false, that a code change is wrong, or that a model actually reasoned from retrieved context. A successful agent session no longer accepts a new baseline automatically: replacing an existing baseline requires explicit review of the selected grounding.

Agents are both readers and contributors: they can retrieve the team's existing context, help capture discoveries from real work, and prepare knowledge proposals or handoffs for a person to review. They do not independently decide what should be published or shared.

Setup installs small host-agent instructions that point to .mex/AGENTS.md for policy and .mex/ROUTER.md for task-relevant context. Agents can query Wiki and Graph evidence, provided the host follows those instructions.

For an existing setup missing your Claude Code or Codex assets, preview and sync that integration explicitly:

mex skills sync --dry-run --tool codex
mex skills sync --tool codex

Use --tool claude for Claude Code. Review the resulting instruction and skill files, commit and push them if the team should share the integration, and start a new agent session.

Instructions may select Inbox or Relay from clear natural-language intent, but skill activation never approves a canonical write. When MEX context materially helps, the agent mentions MEX and the relevant finding naturally, tying it to what it helped understand, decide, or verify, without a standard footer.

With the 0.8.1 Inbox update, “use MEX Inbox to capture what we decided” produces a contribution to existing project knowledge. The agent checks existing records, then drafts one addition or correction. Architecture, components, conventions, decisions, patterns, and guides are supported alongside existing Spec proposals. A local draft stays in the checkout; publishing writes a Markdown proposal for review; approval updates canonical knowledge and retains the proposal as history. Git shares those files with teammates.

Ordinary GROW, Wiki, and context upkeep remains available without Inbox. Inbox is the explicit contribution path, not another knowledge category.

For an existing MEX project, you can manage the official Inbox and Relay skills with skills.sh instead of MEX's skill installer:

npx skills add mex-memory/mex

This installs agent skills only; it does not install the mex CLI or initialize project memory. The skills still need an available MEX CLI and a configured MEX project.

Choose one installer for each agent integration. If mex setup or mex skills sync already manages its skills, keep using mex skills sync; do not run skills add over them. The external installer replaces MEX's managed directories, causing later setup or sync operations to report ownership or path conflicts.

Team workflows use signed previews to bind reviewed inputs and detect stale or altered plans; Wiki authoring uses a plan/--apply boundary. These protect mutation integrity—not authentication, OS isolation, repository permissions, or proof that a human issued the command.

These workflows help a team decide what becomes durable knowledge and preserve enough context for someone else to continue. They sit alongside your existing code-review and issue-tracking tools.

Members provide attribution and provenance. They are not accounts, authentication, role-based access control, or repository permissions.

In 0.8.1, the Inbox skill captures one durable conclusion from a discussion as an addition or correction to architecture, components, conventions, decisions, patterns, or guides. It searches existing knowledge first, saves a local draft, and publishes a Markdown proposal for review. Approval changes the existing knowledge area and retains the proposal's source evidence. Existing spec.create and spec.update proposals remain supported.

Every canonical proposal transition still needs ordinary commit/push/pull to reach another checkout. Approval, rejection, and withdrawal are terminal; a stale proposal can be repaired back to pending. An author can use the exceptional self-approval flow, so Inbox is designed for explicit approval—not guaranteed peer review.

Ordinary GROW upkeep can still update project knowledge directly. Inbox is the explicit contribution-and-review path; session notes remain in the event log.

A Relay packages what the next person needs: a summary, progress, blockers, next actions, useful evidence, and observed repository state. The 0.8.1 update adds Open to team, including active Members who join later, alongside named recipients. Save a local draft before selecting anyone, or use mex relay draft save --from draft.json --json for the shorter agent path. Publication requires an active sender; named handoffs also require one to 32 unique active recipients. The observed state includes branch, HEAD, a dirty-tree flag, and timestamp; it stores no diff or dirty file contents.

A Relay is a durable handoff, not chat, a live notification, task assignment, or a Jira replacement.

Within one observed repository state, the first successful eligible Member becomes the sole claimant. There is no cross-clone network lock, so two unsynchronized Members can claim separately and later meet a Git conflict. Only the active recorded sender or active recorded claimant can close the Relay; deactivating either principal can block closure. In 0.8.1, reactivation restores the original Member identity so older handoffs remain usable. There is no decline, reassign, unclaim, reopen, or administrative-override flow. Teammates exchanging new open-to-team Relays need a CLI that supports schema v4.

mex log records decisions, notes, risks, and todos in .mex/events/decisions.jsonl. Future agents can retrieve relevant entries with mex timeline; people can read them as Project notes in Hub Activity. These records are historical context. Durable conclusions can be promoted into maintained knowledge explicitly, with their source retained.

In 0.8.1, Settings → Agent logging controls when agents write optional notes in this checkout:

• Significant events (default): decisions, discoveries, and risks worth remembering.
• Task checkpoints: batch useful notes when a task or session ends.
• Only when asked: write optional notes on an explicit user request.

The preference guides agents; explicit user requests and mandatory workflow Activity remain independent. It lives under .mex/local/ and is not shared through Git. Start a new agent session after syncing the updated instructions.

mex logging --json
mex logging checkpoints
mex timeline --query "retry" --file src/client.ts --type decision --limit 20 --json

Timeline retrieval reads the latest 8 MiB / 10,000 log lines, returns at most 200 entries, and caps output at 64 KiB. JSON reports truncation; a filtered empty result does not prove the full history has no match. Retrieval never changes the log or initializes project identity.

Run mex <command> --help for the complete interface.

Two commands cover it:

• mex check runs every deterministic check, with no AI and no writes: scaffold files, paths and commands, and every code link in the scaffold and the Wiki. Each link gets one verdict, the same one wiki query, wiki for-code and the Hub show: fresh, changed-nearby (the change is away from everything the fact names), moved, changed, missing or ambiguous. Comment-only and formatting-only edits stay fresh. A finding on a Wiki entity names its id and title.
• mex sync fixes what check found in one interactive agent session. For each flagged Wiki entity the agent compares the fact with the code as recorded and as it is now, then re-records it with mex wiki reground <id> --apply (still true), fixes the text and re-records (changed), or leaves the flag (unsure). Missing and ambiguous links are relinked by hand or left flagged. Sync then rebuilds the Wiki index and checks again.
• mex wiki link-sections previews missing connections from existing section entities to their enclosing document or section. Add --apply to record those refines relationships without changing facts or code groundings. It works on existing scaffolds without setup or an AI session. Semantic dependencies still need evidence. The Hub graph remains bounded to 100 entities and 500 relationships and keeps a limited number of linked parents visible; use List for the complete knowledge collection.
• mex wiki upgrade previews a bounded review of existing broad code-linked entities for one fact per entity. --apply starts an interactive review using the same agent chooser as mex sync; --tool optionally overrides it. Without an installed CLI, copy the preview prompt into your coding agent. Review requires a fresh code graph and refuses flagged candidate groundings until check/sync has reviewed them. It preserves existing ids as overviews and does not rerun setup. Opening Hub or updating MEX never rewrites project knowledge automatically.

mex wiki query and mex wiki for-code are how agents find that knowledge; --health changed lists what check flags.

Use mex capabilities --json for machine-readable capability discovery and mex commands for the concise CLI map.

For a global installation, upgrade the CLI and refresh the selected Claude Code/Codex skill copies:

npm install -g mex-agent@0.8.4
mex skills sync --dry-run
mex skills sync

For an already completed 0.8.0–0.8.3 setup, this refreshes the managed skills and agent guidance; setup does not need to run again just for the package upgrade. Review any reported conflicts with locally edited instructions, then start a new agent session.

The Graph schema remains v4; the disposable Wiki index advances to v5, so run mex wiki rebuild-index after upgrading. The first graph refresh re-extracts the corpus once for the new extraction cache and extractor/resolver versions. To convert existing fingerprints together, preview mex graph compact-fingerprints --dry-run, then run mex graph compact-fingerprints and review the diff before committing. See Upgrading to 0.8.4 for the full sequence.

Keep your existing graph when upgrading from 0.8.2 or earlier. The TypeScript identity changes introduced in 0.8.3 can change node IDs. Refresh before syncing so MEX can map the old IDs to their replacements:

mex graph refresh
mex sync

Review ambiguous or missing groundings and commit the scaffold changes through Git. Rebuild the Wiki index with mex wiki rebuild-index if those changes made it stale. A fresh clone or a rebuild without the old graph relies on fingerprint reconciliation and may require manual re-grounding. Follow the compatibility guide for damaged or incompatible stores; ordinary reads never repair them.

Package and skill upgrades alone do not make an older or incomplete repository Hub-ready. For those repositories, evaluate setup with a dry run before applying it; setup preserves authored files and still requires review of its changes:

mex setup --dry-run
mex setup
git status --short
mex capabilities --json

Review every generated change before committing. In particular, confirm that .mex/graph.db*, .mex/wiki.db*, and .mex/local/ are ignored.

Existing Markdown scaffolds remain valid, and Graph reads never migrate a store implicitly. Compatible schema-v2 and complete schema-v3 stores can upgrade through explicit repair; schema-v1, partial, ambiguous, malformed, or corrupt stores require rebuild. Follow the exact action from mex graph status. Do not add a broad .mex/ ignore rule—it would hide the canonical memory your team is meant to share.

MEX does not upload its canonical records, Graph, Wiki index, drafts, identity selection, or Hub sessions to a MEX service. It provides no automatic team transport: sharing happens through normal Git actions you perform. The Hub binds to loopback, and MEX's local retrieval layer requires no model credentials.

MEX has pseudonymous CLI and Hub usage telemetry, enabled by default unless you opt out. Namespaced commands, outcomes, explicit Hub actions, page categories, and job results help measure feature use and returning installations. Both surfaces reuse one random installation UUID. When available, events also include the existing scaffold UUID to estimate shared-project use and configured AI-tool names from project setup; these do not identify the currently invoking agent or prove team size. Someone with access to project configuration can associate its scaffold UUID with that project. Names, repository remotes, argument values, paths, content, search text, and contact details stay excluded. The ingestion service can observe ordinary transport metadata. Delivery uses a bounded local queue and short CLI cleanup grace; offline telemetry does not change command results.

Check or disable telemetry with:

mex telemetry inspect
mex telemetry status
mex telemetry disable

It can also be disabled with MEX_TELEMETRY=0 or DO_NOT_TRACK=1. See the telemetry policy for the controls, event catalog, privacy boundary, and delivery limits. Coding agents connected to MEX may have their own network and telemetry behavior; that is governed by those tools, not by MEX.

MEX 0.8 does not provide:

• a cloud-hosted Hub or hosted knowledge sync;
• live notifications, presence, or real-time chat;
• Git staging or commits without explicit review, or any pushes or pulls;
• authentication, repository authorization, or RBAC;
• Jira-style task management;
• a shared Code Graph or Wiki SQLite database;
• automatic proof that a model used retrieved context correctly;
• a claimed semantic or vector-search engine—Wiki search is full-text and Graph retrieval is lexical/structural;
• general-purpose Wiki editing through the Hub;
• a published MCP server or MCP package—the source workspace is not a 0.8 public product surface;
• features that appear only in future plans or design documents.

MEX keeps team memory in repository files and provides local retrieval and review workflows. Git and existing engineering tools handle distribution, access, and code review.

• Read the MEX 0.8.4 release notes.
• Check the runtime and compatibility guide and security policy.
• Review the Code Graph support matrix.
• See the extractor model and supported relationships.
• Read the graph retrieval benchmark results, including the blind-graded comparison against an ordinary file-search baseline.
• Inspect the CLI locally with mex capabilities --json and mex commands.
• Join the MEX community on Discord or visit mexmemory.com.