Build1 publisher3 min readPublished Updated
A docs audit splits agent memory into four capabilities the five coding tools answer differently
Vendor pages for Claude Code, Cursor, Codex, VS Code and Windsurf, re-checked on 2026-09-12, show that all five resume a previous session. The complaint about agents forgetting points at the other three questions.
The Engineer · Build desk
What happened
- The published vendor documentation for Claude Code, Cursor, Codex, VS Code and Windsurf was re-checked on 2026-09-12 for a dev.to write-up on why coding agents lose conversation history between sessions.
- Every one of the five already ships a working, documented way to reopen the previous conversation.
- The write-up splits memory into four questions: returning to a conversation, what degrades when you do, what a new session reloads from disk, and whether the agent writes memory for itself.
- Claude Code's documentation states that each session begins with a fresh context window, and on all five the earlier conversation stops at the session boundary.
- All five document compaction, and Claude Code is the only one that documents asking the user before it summarises a session.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- decision A team that has not said which of the four capabilities it needs will compare tools on the one where all five are equal. If the requirement is memory the agent writes itself, resume parity settles nothing.
- constraint With the session as the unit of persistence, anything that has to survive until Monday belongs in a file on disk, and a multi-day workflow has to be designed to reload from files rather than from the thread.
- exposure Teams running long sessions on Cursor, VS Code or Windsurf cannot work out from the published pages when compaction will fire or which tool outputs go, so the loss arrives in the middle of a task.
- precedent Claude Code publishing both a trigger and an itemised list of what a summary destroys gives buyers a specific thing to demand from the other four vendors.
Resume gets the conversation back, and the rest of the setup you supply again. Claude Code documents five entry points, among them a flag for the most recent session in the current directory and a picker for older ones. It says a resumed session restores the conversation together with the state saved in it [4]. Its pages also name what you have to supply again: flags for MCP configuration, settings and added directories [5].
The surface differs by tool. Cursor resumes only from its CLI, and its documentation says the most recent conversation comes back through agent resume, --continue, or the /resume slash command [6], with a session list to pick from as well [7]. Codex resumes the last session or a chosen one, and its own reference marks the resume commands stable [8]. VS Code needs no flag; its agent sessions view restores previous sessions in the interface [9]. Windsurf, now Devin Desktop, resumes from its CLI and from the history panel in the desktop app [10].
What a brand new session does load is the static instruction file a person wrote, and all five reload one [12]. VS Code calls these always-on instructions and includes them in every chat request [13]. Cursor reads its own rules format and other tools' files as well [14]. Windsurf's default agent injects rules at the start of every session [15]. According to the write-up, an instruction file is a thing a person decided once, while memory is what the agent accumulated from what actually happened. So a fresh session that knows your conventions and nothing about last week is an instruction file working with no memory in play [16].
Compaction is where the five stop agreeing [17]. On Pro and Max plans, Claude Code opens a dialog when you resume a session that has been idle for more than about an hour and holds over a hundred thousand tokens. The dialog offers resume from summary, resume the full session as is, or stop asking [18]. It also itemises the loss. The summary replaces the verbatim conversation, full tool outputs and intermediate reasoning are gone, and the instruction file, the auto memory and up to five recently modified files are re-read from disk [19]. Codex compacts on command and automatically at a configurable token limit that ships unset [20]. VS Code says "important details from early in a long conversation might be compressed or lost", and its current pages give no number [21]. Cursor documents a fixed context window shared by every chat and an agent that compacts as it fills; the threshold and the list of what goes are both absent [22]. Windsurf documents a background compaction and a forced one, and stops short of what a compaction keeps [23].
Count the tools whose docs give a trigger you could plan a long session against and you get one of five: Claude Code's hour of idle time and hundred thousand tokens. Codex's automatic limit arrives unset out of the box, and the other three print no threshold [1].
On the fourth question, whether the agent writes memory for itself, the audit's documented example is Claude Code's auto memory, on by default in its own words [24]. That file is also one of the things read back off disk after a summary resume [19].
All of this is a documentation audit dated 2026-09-12 [1]. It transfers to your machine only if the build you installed matches the page the vendor published, and a resume flag Codex's own reference calls stable is a safer bet on that than an undocumented compaction trigger [8].
What to watch
- Whether Codex ships its automatic compaction limit with a default value instead of leaving it unset.
- Whether Cursor, VS Code or Windsurf publish a compaction threshold or an inventory of what a summary drops.
- Whether any of the other four document an agent-written memory of their own, since the audit's example is Claude Code's auto memory.