# recall 0.5.0

Released 16 Oct 2026 · [Compare on GitHub](https://github.com/alirng/recall/compare/v0.4.2...v0.5.0) · [recall-agents 0.5.0 on PyPI](https://pypi.org/project/recall-agents/0.5.0/)

## Upgrading

- **Codex now asks before `message_chat` runs.** `message_chat` writes a message to another chat's mailbox, and Codex's default `auto` approval let it through without asking, because the tool is marked neither destructive nor open-world. Run `recall install` to add the approval to your existing Codex entry (`recall doctor` says when an entry lacks it); an approval you set yourself is left alone, and `recall uninstall` removes it with the entry.
- **The chat index rebuilds itself once more.** The first index run after upgrading re-reads your chats, and Cursor's database in a background run after that. On Linux, Cursor chats that had failed to index because Linux keeps no file creation time come back.
- **The notes folder's git history still holds the start of every note's body.** Earlier releases kept it in the deletion ledger, `.recall/notes.json`, which each memory run commits. The first memory run on this release rewrites the ledger without bodies; once it has (no note text left in that file), this removes the old versions from every commit, needs git-filter-repo, and removes the folder's git remote, so add it back if you sync the notes: `cd ~/.agents/memory && git add -A && git commit -q --allow-empty -m "Before purging the ledger's history" && cp .recall/notes.json ../recall/notes-ledger.json && git filter-repo --force --invert-paths --path .recall/notes.json && mkdir -p .recall && mv ../recall/notes-ledger.json .recall/notes.json && git add -A && git commit -qm "Deletion ledger without note bodies"`.

## Changed

- **recall waits for power on battery.** `recall dream` defers itself when it starts on battery outside a terminal, the nightly run included (launchd runs a missed night when the Mac wakes), and the first tool call or session start on power runs it in the background. Tool calls start no background indexing or embedding on battery, and the jobs recall starts in the background run at low priority. `recall status` and `recall doctor` say when a run was deferred. `recall dream --on-battery run` runs one anyway, and `on_battery = "run"` under `[power]` in config.toml lets everything run on battery.
- **recall's hooks and tools answer faster.** On a history of 1,200 chats, the edit hook took 80 ms, 175 ms while other chats were working, and about 0.8 s with its cache gone, because before every edit it listed every agent's transcripts and read again in full each chat that had moved in the last 12 hours. Now it reads only what each recent transcript gained since its last look, and reads chats in full and asks git only when another chat touched the same file: 40 ms, and 150 ms cold. A chat that starts without recall's hooks (Codex before you trust them) is still seen at the next edit, from the folder its agent writes it to. One resumed in a transcript older than a day is seen within five minutes. When another chat has the file open, the git calls run side by side and skip macOS's `/usr/bin/git` stub (380 ms down to 150 ms). A prompt no longer waits on an Ollama model that is not loaded. Asking one started a load that Ollama dropped when the hook gave up, so every prompt waited a quarter of a second and the model never loaded. Now a background process loads it, at most once every 25 minutes, and a loaded model is waited for at most 100 ms rather than 250. Hooks that do not read the index load 15 ms less code. The MCP server reads the shell's rc files once per index refresh rather than three times, memory notes load eight times faster, and a tool call waits for an index refresh only until it is a second old. A longer one, such as the rebuild after an update, carries on in the background. Until it is done, `search_chats`, `recent_chats` and `read_chat` say that chats may be missing. Refreshes commit after each chat they read, and the redaction scrub every half second, so a tool call that writes no longer waits tens of seconds for the index. One that would wait over 3 s answers from the notes or commit links as they stand.
- **Semantic search keeps one copy of the vectors, not one per agent.** Each agent session runs its own recall MCP server, and with embeddings on, each built a matrix of every message vector on its first search and kept it: over 100 MB a server for 23k messages, more than 1 GB with ten agents open. The vectors now sit in one file beside the index (`recall-vectors/`) that a server maps only while it searches, so the operating system holds one copy for all of them. Five servers after two searches each take 428 MB instead of 830 MB, and search ranks as before: search_v2 and SWE-CB were identical on a same-day paired run. `recall forget` rewrites the file without the forgotten chats. `recall embed` no longer runs twice at once and stops after 20 minutes (the next run carries on), finding the messages to embed reads only the chats that changed, and a search waits at most 10 seconds for the embedding model before going by words.

## Fixed

- **One chat recall cannot index no longer breaks its tools.** A chat that failed to index made every tool call fail and kept the chats after it out of the index, and a damaged archive copy did the same. Such a chat is now logged by its path, never its words, and tried again after a minute, ten minutes and an hour, then left until its file changes, while the others index as usual. `recall doctor` lists these chats and `recall stats` counts them. A damaged archive copy is replaced by the next read of its file, and can no longer stop recall from opening the index after an update.
- **Tool calls no longer rescan Cursor's database every 30 minutes.** Cursor's older chats, and its subagent and side chats, sit in a database of several GB. recall now reads it in full in the nightly run and on `recall index`, and from a tool call only on the first index or a day after the last full read (in the background, at low priority, stopped after 15 minutes), and skips it while the database is unchanged. Other chats are still read as soon as they change. A new Cursor side chat becomes searchable after the next nightly run or `recall index`.
- **Memory that reaches an agent unasked no longer carries commands, links or other chats' prompts.** A README or web page an agent read could become a note that session start injected into every later session. Now a rule that gives an agent something to act on (a URL or host, a command to run, a check switched off, private data sent somewhere) never loads at session start or mid-session, and neither does a note with such a title; a prohibition such as "never force-push main" still loads, and memory_search and read_note still find everything. A note whose rules only the agent stated is counted, not named, and rules a user turn stated without a typed quote are now marked `unverified`, so an unmarked rule means one from before speakers were recorded. Injected text keeps no `<` or `>` (each becomes ‹ or ›), so no note, rule, past chat or exploration report can open or close the block, and titles are one line of at most 120 characters. On my chats, session start loads 0.574 of the notes it must (was 0.576), 0.482 on held-out folders (was 0.488), holding back 13 of 733 rules. `recall status` redacts a chat's title and latest request before caching them or quoting them to other chats, and its cache rebuilds once. The deletion ledger (`.recall/notes.json`) kept the start of every note's body, and consolidation sent 200 characters of each deleted note to the model: it now keeps the title, scope and project, existing ledgers are rewritten without bodies, and a deleted note restored on disk leaves it. The history rewrite `recall forget` prints now commits the notes folder first (filter-repo's reset would discard uncommitted hand edits), deletes the listed notes from disk and history, also removes the ledger's past versions and puts the current one back, and passes `--force`, without which filter-repo refused to run.
- **memory_search no longer rebuilds the notes index on every call.** It used to drop and rebuild the notes tables and load every note twice per search, which took the index's write lock while other agents' servers were reading, so a search elsewhere could fail with a transient `vtable constructor failed` error. It now rebuilds only when a note file, the `[contexts]` in config.toml, or the tables' layout differs from what the index was built from, so a hand edit still shows up on the next search, and a rebuild is one transaction. With 733 notes and embeddings on, a search takes about 35 ms through the MCP server, against about 530 ms in 0.4.2. Servers still running an older release keep rebuilding the tables until their agent restarts.
- **A model server on another machine is no longer called private.** A local model whose `base_url` is not `localhost` or a loopback address used to be reported as "nothing leaves this machine". `recall dream` no longer says so, and setup names the host the chat excerpts go to.
- **SECURITY.md and the README say what recall does.** They said all six MCP tools only read. There are nine, and `message_chat` writes to recall's mailbox. SECURITY.md now also lists every file `recall install` writes, what the hooks can do (refuse an edit once, keep an agent going, and rewrite a subagent's prompt if you turn that on), what a hosted embedding endpoint receives, and that recall keeps its copy of a chat after an agent deletes it, even past a short `cleanupPeriodDays` you set yourself.
- **Time windows as agents write them.** `since` and `until` accept `2mo`, `3 months`, `2wk`, `5 days`, `4 hours`, `1yr` and the like, not just `12h`, `3d`, `2w`, `1m` or `1y`. A window recall cannot read now says which forms it takes, instead of failing with a bare date error.