Changelog

• Claude Code reports a failing Bash command as PostToolUseFailure, not PostToolUse. observe.sh now listens there too, so a failing command is recorded when the mod is not running (older Claude Code, mods off, or a mod that failed to load)
• An interrupted command (Esc) is not recorded
• Standalone installs register the new hook on their next session start
• Secrets in a failing command or its output (API keys and tokens, Authorization and Bearer values, passwords in URLs, token= and password= values and flags, private keys) become [REDACTED] before the memory leaves your machine. Commit SHAs stay

• When a newer Recall is out, the row now says to refresh the marketplace first: /plugin marketplace update recall-claude-plugin, then /plugin update recall
• Tip: turn on auto-update for recall-claude-plugin in /plugin → Marketplaces, and new versions arrive on their own

• In a narrow window a long warning in Recall's row no longer pushes the band's last row out of view

• Recall's workspace now starts at the same column as the values in other products' rows
• One blank row above the band when it has room; in a short window the rows come first

• A status line set up before the plugin now updates with it, so it stops showing Recall where Recall already has its own line above the prompt

• The status line leaves Recall's segment out only when Claude Code really loads Recall's mod, so a curl install always keeps its segment
• Recall's row above the prompt starts 2 cells in, lined up with the status line

In 1.18.0 the status line dropped its Recall segment only once the mod had started. The status line draws before that and does not redraw while Claude Code is idle, so a session could show Recall in both places. The status line now decides from what is installed.

• Claude Code 2.1.287 or later, mods on, and the recall plugin at 1.18 or later: Recall shows only above the prompt
• Older Claude Code, VS Code's chat panel, claude -p, or the curl install without the plugin: the status line segment works as before
• Recall's line is one row: Recall in bold, the workspace, then dim detail. It no longer says "workspace not confirmed", and turns yellow only when something failed

The plugin and its scripts update by themselves. To get it at once:

claude plugin marketplace update recall-claude-plugin
claude plugin update recall@recall-claude-plugin

Recall used to add a segment to your status line, wrapped around whatever status line you had. It now has a line of its own in the band above the prompt, and your status line is yours again.

🧠 Recall 1.18.0 · 2 stored · workspace confirmed · stored a memory (5s ago)

• Version, memories stored this session, workspace, recent activity, and a notice when an update is out
• Lines that other plugins draw above the prompt stay
• Where mods do not draw (VS Code's chat panel, claude -p, older Claude Code), the status line segment works as before

Recall remembers a shell command that failed, with its output. The plugin used to start a script for that after every shell command, about 290 ms and ten small programs each time. The mod does it inside Claude Code and sends what it found every 15 seconds, so the command's result reaches Claude at once.

When a session's workspace has expired, Recall refuses to store into it (WORKSPACE_NOT_CONFIRMED). Claude had to remember to confirm the workspace and try again. The mod now does it: it confirms the workspace and runs the call once more. It never loops.

The mod comes with the plugin, not with the background script update. Update the plugin, then restart Claude Code:

claude plugin marketplace update recall-claude-plugin
claude plugin update recall@recall-claude-plugin

Or in Claude Code: /plugin, Installed, recall, Update, then /reload-plugins. It needs Claude Code 2.1.287 or later (claude --version).

The hooks used the shell's current folder, which follows every cd. A session that stepped into a subfolder, a git worktree, a scratch folder or /tmp started a new workspace there, and memories saved there never showed up in the project. One account had 217 workspaces this way.

• The workspace is the folder Claude Code was started in, for the whole session
• Resolved to its git checkout: the repo root for a subfolder, the main checkout for a worktree
• An explicit RECALL_WORKSPACE still wins

The Workspaces page now groups workspaces by the folder that holds them and suggests what to tidy: worktrees, scratch folders and subfolders to link into their project, and empty workspaces to delete. One click each, nothing without it.

• Linking keeps a workspace's memories searchable from its project, and new ones go to the project
• Linked workspaces don't count toward your workspace limit, and can be unlinked any time
• Sibling projects under one folder are left alone

A repository cloned over https with a token in its URL keeps that token in its remote. Recall used the remote to identify the workspace and sent it as it was. The hooks now remove the user and token first, and the server removes them again before storing or hashing, so older clients are covered and a rotated token keeps the same workspace.

The installer used to publish Recall on every network interface, and Docker's published ports bypass host firewalls. A new install now listens on 127.0.0.1. Set RECALL_BIND=0.0.0.0 (behind a TLS reverse proxy) to reach it from other machines; a reinstall keeps the address it already had.

Memory fetches are now pipelined in batches instead of one round trip per record, and decrypted embedding vectors are cached in-process. Semantic searches that previously took tens of seconds on large workspaces now return in well under a second, with the relevance floor keeping unrelated results out.

• Batched memory fetches across workspace, linked and global scopes
• In-process vector cache (RECALL_VECTOR_CACHE_SIZE, default 10,000)
• Same relevance floor: gibberish returns nothing, not noise

Per-call capture of every file read, search and edit produced 99% of stored memories while being retrieved essentially never - and crowded real memories out of search results. The observe hook now stores only failing commands, with the error output attached so the memory is actually useful later.

• observe.sh captures only failing Bash commands, with an output excerpt
• One session summary per session (SessionEnd), no longer one per turn
• Summaries no longer carry embeddings they never needed

• Double-registered hooks no longer store every observation twice - registration is reconciled automatically, and a server-side guard collapses identical rapid writes into one record
• The pinned-workspace write guard now applies to REST exactly as it does to MCP - a restored session cannot silently file memories under a guessed project
• set_workspace on the workspace you already hold correctly lifts the write refusal after a session expires

Restarting into a resumed session never triggered the updater, so "Restart Claude Code to update" could nag forever. The update check now runs on resume as well, and the self-updater keeps every installed copy - including the one your status line reads - on the same version.

And in v1.17.2: sessions now stay writable for 4 hours of idle time instead of 30 minutes, and when a genuinely stale session does need re-confirming, agents recover silently with a re-confirm and retry instead of surfacing an error.