hotbotspot.com ยท bot technologies

Coding workers and build-your-own foundations

As of Oct 25, 2026. Documentation-based, not a hands-on benchmark.

These agents mostly work on software projects. Their persistent instructions, files and memory overlap a lot with the continuing-assistant idea. Having memory doesn't make an agent an always-running project coordinator, though.

ProductWhere memory livesWhat it leaves out
Cursor AutomationsNamed memory entries (MEMORIES.md by default), stored outside the working filesystem and shared across runs of the same automationNot shared across every Cursor agent; it belongs to that one automation
Claude CodeCLAUDE.md instruction files, plus an auto-memory directory per project with a MEMORY.md index; it stays on one machineAt startup it loads only the first 200 lines or 25 KB of the index; other topic files are read when needed
Grok BuildMarkdown notes, consolidated into topic files by /dreamTask state, tentative conclusions, secrets, and anything the repository or its documentation already covers
OpenCodeAGENTS.md instructions and saved sessions; long-term memory only through pluginsAutomatically curated long-term memory, unless you add a plugin

Sources: cursor, code.claude, x, projectmem, github

Cursor Automations and Cloud Agents

What the documentation says. Automations run on a schedule or on events from services such as GitHub, GitLab, Slack, Linear and webhooks. Runs of the same automation share persistent memories. Memories are named entries, with MEMORIES.md as the default. They're stored outside the working filesystem, so they outlast any one agent's workspace. You can view, edit and delete them in the configuration screen. cursor

The memory belongs to one automation. It isn't shared across every Cursor agent. Cursor also warns that untrusted input can create misleading or malicious memories that affect later runs. cursor

Cloud Agents are remote agents that work in isolated development environments while you do other things. They're best understood alongside Automations, not as a separate bot platform. tembo, stevekinney

Our assessment. This is very close to "write down what you learned so you do the job better next time." For example, a recurring review agent can remember that it already raised an issue. When a bot learns from its work automatically, you need controls over what becomes a standing instruction.

Claude Code

What the documentation says. Claude Code has two ways of keeping things. CLAUDE.md files hold the instructions you set on purpose, for a project, your own workflow or your organization. Auto memory is the notes Claude writes for itself across sessions: corrections, preferences, project decisions and references. It writes them to a project memory directory, with MEMORY.md as an index. Subagents can keep their own auto memory. At startup it loads the first 200 lines of the index or 25 KB, whichever comes first, and reads other topic files when it needs them. Auto memory stays on one machine, and worktrees in the same repository share it. code.claude

Our assessment. It is a file-backed way of carrying selected learning forward, not just a compressed chat. For software projects it may already cover much of what we want from a bot. Unlike OpenClaw, it is built around doing project work, not around standing scheduled duties.

Grok Build

What the documentation says. After a turn is finished, Grok Build reviews the work in the background. It saves durable conventions, decisions and project facts as Markdown notes. Its /dream process consolidates those notes into topic files. Before related work, it reads the relevant topics, even when you haven't brought them up. You can inspect memory with /memory and fix a wrong note in its file. Instructions in the current conversation take precedence over stored notes. The memory deliberately leaves out task state, tentative conclusions, secrets, and anything the repository or its documentation already covers. x

Our assessment. "Remembers how we work" isn't the same as "keeps the official list of unfinished commitments." Grok Build is a coding worker. It doesn't do Grok Bot's continuing-assistant job.

OpenCode

What the documentation says. OpenCode keeps persistent project instructions in AGENTS.md. It saves sessions that you can resume, fork or export. projectmem, supermemory Plugins fill the gap between a saved session and real memory. The opencode-agent-memory plugin adds memory blocks with project and global scopes, memory tools, and prompts that push the agent to maintain its memory. github The Supermemory integration adds saving and recall across sessions. supermemory OpenCode also compacts context, so long-term memory doesn't remove the need to summarize the current session. iceberglakehouse

Our assessment. "I can reopen an old conversation" isn't the same as "the agent will spot the relevant lesson in a new task." Compare a specific OpenCode setup, including its plugins. Don't assume the base product behaves like a continuing assistant.

Sources