Skip to main content
Open the worker URL printed by claude-mem, usually http://localhost:37777. My Work is the default local viewer. It brings recorded human requests, progress, next steps, and finished results into one readable place. The panel runs locally and reads your saved memory database. It does not require a cloud account, a new model call, or an internet connection to show your work. The original memory timeline, sessions, settings and console remain under Memory & settings, at /viewer on the same worker. Opening the panel and switching between its pages are the only things it reports: two anonymous telemetry counts (work_view_opened, work_view_switched) that carry the page name and nothing else. They never include your requests, prompts, observations or file paths, and npx claude-mem telemetry disable turns the counts off.

​
Browser setup and your prompts

The browser setup reads your local agent histories (last 30 days) on this computer only. Nothing is uploaded until the scan finishes and you choose Discover My Work, Rules & Focus. That upload sends the full text of the prompts you typed, with session IDs, project names, folder paths and timestamps, to cmem.ai, which processes them with a third-party AI model provider. You can untick any tool before uploading, or choose Skip this step to send nothing. The separate Share anonymous usage statistics to help improve your Claude-mem experience box is checked by default. You can uncheck it before continuing; it does not start transcript analysis.

​
Working, ready and checked

  • In progress means the latest saved work state says work is underway. It is not a live-presence signal or a promise that an agent is currently running.
  • Needs you appears when recorded work explicitly asks for human input. Open the handoff, inspect the original request and evidence, then continue in the agent conversation where the work is happening.
  • Ready to check means the parent request was explicitly reported complete. A finished subtask or a closed agent session does not finish its parent.
  • Checked means you pressed Mark checked for that version of the result. It records your review locally; it does not certify correctness or change the agent’s source status. Undo is available on a checked card.
  • No recent update means no newer activity has been recorded for six hours.
A new substantive source update invalidates an older check. If the request is still reported complete, it returns to Ready to check. If it has reopened, the recorded work status is shown instead. Review notes live in the private work-view-reviews.db sidecar in the claude-mem data directory and are not uploaded by cloud sync.

​
Live updates

The panel shows new local data as soon as it arrives. Applied cards preserve their existing order. Timestamps refresh without rebuilding the card you are reading. Local updates connected describes the local data connection, separately from cloud sync. If the connection or source fails, the panel keeps its last received work visible and shows the interruption. It never animates a saved work status as proof of agent activity.

​
Requests and source evidence

Open a request to see its original human ask, recorded follow-ups, result and source records. Copy handoff copies the goal, recorded state, next step and recent updates so you can continue in the agent you choose. It never sends or assigns work; if clipboard access is unavailable, select the displayed text. A captured conversation normally forms one request; explicit request_id fields can distinguish independent goals. The panel does not guess that two unrelated conversations describe the same request. Records without a proven human request remain under work records need a linked human request. The live view loads a bounded recent slice; use Memory & settings for the full saved history. Search and the time filter scope the cards, state counts and Your next move together. When several projects are loaded, use the project selector to narrow that same view. Other states includes blocked, queued, quiet and unknown work; each card retains its recorded state label.

​
Deliverables

Ask for a deck, a report, an HTML page, an image or any other file, and it is one click away on the front of the request’s card under Deliverables, with the full list in the request drawer under What came out of it. When an agent stops, claude-mem’s Stop hook copies what the turn made for you into the project’s deliverables bucket, ~/.claude-mem/deliverables/<project>/, and records each copy. The bucket keeps the file after the place it was made (a worktree, a scratch folder, a temporary directory) is deleted, so the link keeps working. Each copy is named by its content (report-3f9a1c2b.md), so the same file lands once and a revised file lands beside its earlier version. Two things count as a deliverable:
  • Attached by the agent. A work_state_write whose work_product (or deliverable) field names a local file, with any extension, or any other field whose value is a path to a file a person opens (output_pdf=/path/slides.pdf). claude-mem’s MCP server instructions ask agents to write the thing you asked for as a file and attach it before they finish.
  • Written outside the source tree. A deck, document, page, image, video or spreadsheet the agent wrote with its file tools into a git-ignored folder such as .scratch/ or outside any repository. Source files are never copied by inference, so a .html component in src/ stays where it is; an agent attaches it if it is a deliverable.
Markdown and text deliverables open inline; pages and images open sandboxed, the same way recorded files do. While a copy is in the bucket and the file it came from has not been written since, the copy stands in for the record of where the file was made; a file revised after the copy shows beside it as a newer version. A file is copied only from the checkout or the system temp folders, where the host’s scratchpad lives; a path elsewhere, or a symlink that leads elsewhere, is left out and logged, because a work-state field can name any path without a permission prompt. A turn whose prompt is entirely <private> leaves no copy and no record, the same rule observations and summaries follow. The bucket is local to this computer and is not uploaded by cloud sync; when the worker runs on another machine (CLAUDE_MEM_WORKER_HOST is not this computer), nothing is copied, because that worker could not open the copy. CLAUDE_MEM_DELIVERABLES_ENABLED=false turns the copy off; files already in the bucket stay.

​
Optional Claude.ai and ChatGPT connections

The Connect Claude.ai and Connect ChatGPT actions open the existing CMEM Cloud setup. Clicking does not connect an account or upload memory. Sign in and review the setup and sharing scope before enabling anything. Cloud sync can upload saved prompt text and observation narratives. Your local viewer continues to work without it.

​
Work with your team

Switch from My Work to TeamBrain to view shared work. Personal requests and local review notes remain in My Work. TeamBrain reads the shared board for your selected team using your existing sync connection. If this install has not joined a team, choose Set up TeamBrain to open the team setup and plan page. To join with an invitation, run npx claude-mem teambrain join <invite-link>, approve the device, and choose which projects to share. Opening the TeamBrain tab does not enable sharing. The connected panel shows your actual membership, whether you are broadcasting, the shared project’s open lists and tasks, and linked collision or delegation findings. These are shared work lists, not inferred human requests. Local Checked notes do not appear on them. Shared updates wait while you read, and the panel refreshes only while its tab is open and visible. If access is revoked or a team stops, the shared board is cleared. A temporary connection failure keeps the last received board with a visible error. Team workspace & sharing settings opens the full cloud page for membership, invitations, billing and explicit sharing changes.