# Claude-Mem 14: How My Work Tracks and Reviews Agent Tasks

Claude-Mem 14 adds a local My Work view for agent requests, evidence and human review, with optional TeamBrain sharing and separate recorded-status limits.

By Paras Madan | 2026-10-19


Claude-Mem 14, released on October 17, adds My Work as its default local view of agent requests, progress and results. It gives users a place to inspect evidence and record that they checked an output, addressing the gap between an agent reporting completion and a person reviewing the work.

Claude-Mem is an independent, third-party project, not an Anthropic product. Its [14.0.0 announcement](https://github.com/thedotmack/claude-mem/releases/tag/v14.0.0) introduces Cards, List and Agents views, with attention-first sorting and summaries written through a compatible configured model. With no compatible provider configured, the interface uses recorded wording instead of generating a new summary.

## My Work reads the records on your machine

The [My Work guide](https://github.com/thedotmack/claude-mem/blob/v14.0.0/docs/public/usage/my-work.mdx) places the view at the local worker address, usually `localhost:37777`. The existing memory timeline, sessions and settings remain available at `/viewer` on that worker.

Displaying saved records requires no cloud account, new model call or internet connection. That applies to reading the local view. Generating summaries and running the underlying agents still depend on their configured providers and connections.

This separation makes the page useful for resuming work. Someone can inspect what was recorded before deciding whether to reopen an agent, request another explanation or continue from the original conversation. The source request and evidence remain important when a short summary leaves out a constraint.

## Completion and human review have different meanings

“In progress” reflects the latest saved state, rather than proving that an agent is currently running. “Ready to check” requires the parent request to have been reported complete. A completed subtask alone does not close the larger request.

Marking a result “Checked” records a local human review of that version. It does not certify correctness or alter the source agent's status. A substantive source update invalidates the earlier check, so changed work can return for review.

The distinction is useful for tasks with several outputs. A coding agent might finish its edits while the person responsible for the change still needs to examine the diff and test the relevant behavior. A visible review state can help track that handoff, provided the reviewer actually inspects the result.

There is a smaller attribution fix to know about too. [Version 14.0.1](https://github.com/thedotmack/claude-mem/releases/tag/v14.0.1), published later on October 17, preserves the agent attribution on cards when updates arrive. Users installing the 14 series should check which patch they have.

## TeamBrain adds optional shared work

The release also introduces self-serve TeamBrain setup for work a team chooses to share. Opening its tab does not automatically upload projects, and local review notes remain separate from the shared board. Review the selected projects and sharing controls before treating a personal workflow as a team workspace.

The [project README](https://github.com/thedotmack/claude-mem/blob/v14.0.0/README.md) describes the broader installation flow, including optional hosted memory and provider choices. Claude-Mem's core is Apache-2.0 licensed, but that does not make model calls or hosted services free. The local viewer's availability should be evaluated separately from those costs.

TeamBrain's public pricing needs confirmation. The release announces $50 per seat per month, while the [pricing page](https://cmem.ai/pricing) still describes a design-partner pilot without an announced public price. That mismatch leaves the announcement as a quoted offer, rather than a verified checkout total; confirm the plan shown during setup before budgeting a rollout.

## Use the queue to decide what needs attention

A practical first check is to open a known completed request, inspect its supporting records and mark it reviewed. Then check what happens when the source receives another substantive update. This tests the part of the workflow that matters most: whether a person can distinguish finished, changed and still-unreviewed work.

My Work can make that review easier to organize. The quality of the answer, code or handoff still has to be established through the evidence and whatever checks the task requires.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help AI founders explain memory systems, agent workflows and product launches through technical writing and founder content that reaches the right users.

[Book a 30-minute call](https://cal.com/paras-madan/30min).