diff --git a/src/openhuman/memory/tree/jobs/mod.rs b/src/openhuman/memory/tree/jobs/mod.rs
index cdb71ccb0..77e8287f2 100644
--- a/src/openhuman/memory/tree/jobs/mod.rs
+++ b/src/openhuman/memory/tree/jobs/mod.rs
@@ -26,6 +26,7 @@
 //! See [`store::enqueue_tx`] for the in-tx producer entry point.
 
 mod handlers;
+mod redact;
 pub mod scheduler;
 pub mod store;
 pub mod testing;
diff --git a/src/openhuman/memory/tree/jobs/redact.rs b/src/openhuman/memory/tree/jobs/redact.rs
new file mode 100644
index 000000000..6c0099a68
--- /dev/null
+++ b/src/openhuman/memory/tree/jobs/redact.rs
@@ -0,0 +1,226 @@
+//! Log-side scrubber for free-form `reason` / `last_error` strings emitted
+//! by [`worker::run_once`] (defer/fail branches) and the matching
+//! [`store::mark_failed`] / [`store::mark_deferred`] log lines.
+//!
+//! Persisted DB state (`mem_tree_jobs.last_error`) keeps the full
+//! original string for diagnostics — handlers may attach
+//! upstream-provider responses or full anyhow chains there. Logs are a
+//! lower-trust sink (often forwarded to remote log aggregators or
+//! shared in bug reports), so we apply a uniform scrub policy to:
+//!
+//! 1. Mask credential-shaped tokens (Bearer, OpenAI `sk-…`, GitHub
+//!    `ghp_…`, Slack `xox?-…`, generic `api_key=…` / `password=…` /
+//!    `token=…` assignments).
+//! 2. Strip URL userinfo (`https://user:pass@host` → `https://***@host`).
+//! 3. Mask bare email addresses.
+//! 4. Cap the logged string at [`MAX_LEN`] bytes, suffixed with
+//!    `…(truncated, N more bytes)` so a reader knows the original
+//!    string was longer.
+//!
+//! [`worker::run_once`]: super::worker::run_once
+
+use std::sync::OnceLock;
+
+use regex::Regex;
+
+/// Upper bound on the byte length of a scrubbed string. Long enough to
+/// keep the head of an anyhow `{:#}` chain (which usually includes the
+/// most useful context) but short enough that one bad job can't flood
+/// the log with megabytes of provider response body.
+pub(crate) const MAX_LEN: usize = 1024;
+
+/// Scrub a free-form error / reason string for emission to logs.
+/// Returns an owned `String` because every regex pass may rewrite the
+/// input; callers are emitting this through `format!` / `log::*!`
+/// anyway, so the allocation isn't on a hot path.
+pub(crate) fn scrub_for_log(input: &str) -> String {
+    let mut out = input.to_owned();
+    for (re, replacement) in patterns() {
+        // `Cow::into_owned` is cheap when the regex didn't match.
+        out = re.replace_all(&out, *replacement).into_owned();
+    }
+    truncate(out)
+}
+
+fn truncate(mut s: String) -> String {
+    if s.len() <= MAX_LEN {
+        return s;
+    }
+    // Round down to a char boundary to avoid splitting a multi-byte
+    // UTF-8 sequence — `truncate` itself panics on a non-boundary.
+    let mut cut = MAX_LEN;
+    while cut > 0 && !s.is_char_boundary(cut) {
+        cut -= 1;
+    }
+    let dropped = s.len() - cut;
+    s.truncate(cut);
+    s.push_str(&format!("…(truncated, {dropped} more bytes)"));
+    s
+}
+
+fn patterns() -> &'static [(Regex, &'static str)] {
+    static PATTERNS: OnceLock<Vec<(Regex, &'static str)>> = OnceLock::new();
+    PATTERNS.get_or_init(|| {
+        vec![
+            // URL userinfo: capture the scheme + ://, then the userinfo
+            // up to '@'. Replace with ***@. Anchored on `://` to avoid
+            // eating bare `user@host` strings — those are handled by
+            // the email rule below.
+            (
+                Regex::new(r"(?P<scheme>[a-zA-Z][a-zA-Z0-9+.\-]*://)[^\s/@]+@").unwrap(),
+                "$scheme***@",
+            ),
+            // Bearer tokens (with or without an `Authorization:` prefix).
+            (
+                Regex::new(r"(?i)bearer\s+[A-Za-z0-9._\-+/=]+").unwrap(),
+                "Bearer ***",
+            ),
+            // Provider-prefixed credentials with stable, well-known
+            // shapes. Listed individually so a future reader can see
+            // exactly which providers are covered.
+            (
+                Regex::new(r"sk-[A-Za-z0-9_\-]{16,}").unwrap(),
+                "sk-***",
+            ),
+            (
+                Regex::new(r"ghp_[A-Za-z0-9]{20,}").unwrap(),
+                "ghp_***",
+            ),
+            (
+                Regex::new(r"ghs_[A-Za-z0-9]{20,}").unwrap(),
+                "ghs_***",
+            ),
+            (
+                Regex::new(r"gho_[A-Za-z0-9]{20,}").unwrap(),
+                "gho_***",
+            ),
+            (
+                Regex::new(r"xox[abprs]-[A-Za-z0-9\-]{8,}").unwrap(),
+                "xox-***",
+            ),
+            // Generic `key=value` assignments where the key name implies
+            // a secret. Matches `api_key`, `apiKey`, `api-key`,
+            // `password`, `passwd`, `pwd`, `token`, `secret`. Accepts
+            // a quoted, single-quoted, or bare value; bare values stop
+            // at the first whitespace / comma / closing bracket so we
+            // don't eat the rest of the message.
+            (
+                Regex::new(
+                    r#"(?i)(?P<k>api[_\-]?key|password|passwd|pwd|secret|token|auth)\s*[:=]\s*(?:"[^"]*"|'[^']*'|[^\s,}\)\]]+)"#,
+                )
+                .unwrap(),
+                "$k=***",
+            ),
+            // Bare email addresses. Conservative pattern (no UTF-8
+            // local parts, no quoted local parts) — sufficient to mask
+            // the common cases without eating unrelated `@` symbols.
+            (
+                Regex::new(r"\b[A-Za-z0-9._%+\-]+@[A-Za-z0-9.\-]+\.[A-Za-z]{2,}\b").unwrap(),
+                "***@***",
+            ),
+        ]
+    })
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn passthrough_for_safe_string() {
+        let input = "rate_limited: provider returned 429, retry in 30s";
+        assert_eq!(scrub_for_log(input), input);
+    }
+
+    #[test]
+    fn masks_bearer_token() {
+        let s = scrub_for_log("Authorization: Bearer abc123.def-456_xyz");
+        assert!(s.contains("Bearer ***"), "got {s:?}");
+        assert!(!s.contains("abc123"));
+    }
+
+    #[test]
+    fn masks_openai_key() {
+        let s = scrub_for_log("upstream returned 401: invalid sk-abcDEF1234567890ZZZZ key");
+        assert!(s.contains("sk-***"));
+        assert!(!s.contains("sk-abcDEF1234567890ZZZZ"));
+    }
+
+    #[test]
+    fn masks_github_token_variants() {
+        for raw in [
+            "ghp_abcdefghij1234567890ABCD",
+            "ghs_abcdefghij1234567890ABCD",
+            "gho_abcdefghij1234567890ABCD",
+        ] {
+            let s = scrub_for_log(&format!("error: token {raw} rejected"));
+            assert!(s.contains("***"), "input={raw} out={s}");
+            assert!(!s.contains(raw), "input={raw} out={s}");
+        }
+    }
+
+    #[test]
+    fn masks_slack_token() {
+        let s = scrub_for_log("posting to slack failed: xoxb-1234567890-abcdEFG");
+        assert!(s.contains("xox-***"));
+        assert!(!s.contains("xoxb-1234567890"));
+    }
+
+    #[test]
+    fn masks_generic_secret_assignments() {
+        let inputs = [
+            ("api_key=hunter2 trailing", "api_key=***"),
+            ("password: hunter2", "password=***"),
+            ("Token = hunter2,more", "Token=***"),
+            ("apiKey=\"hunter2\"", "apiKey=***"),
+        ];
+        for (raw, expect) in inputs {
+            let s = scrub_for_log(raw);
+            assert!(s.contains(expect), "input={raw:?} out={s:?}");
+            assert!(!s.contains("hunter2"), "input={raw:?} out={s:?}");
+        }
+    }
+
+    #[test]
+    fn strips_url_userinfo() {
+        let s = scrub_for_log("connect failed https://alice:s3cret@db.internal/x");
+        assert!(s.contains("https://***@db.internal/x"), "got {s:?}");
+        assert!(!s.contains("alice"));
+        assert!(!s.contains("s3cret"));
+    }
+
+    #[test]
+    fn masks_email() {
+        let s = scrub_for_log("user alice@example.com triggered job");
+        assert!(s.contains("***@***"), "got {s:?}");
+        assert!(!s.contains("alice"));
+        assert!(!s.contains("example.com"));
+    }
+
+    #[test]
+    fn truncates_oversized_input() {
+        let big = "x".repeat(MAX_LEN * 2);
+        let s = scrub_for_log(&big);
+        assert!(s.len() < big.len());
+        assert!(s.contains("(truncated,"));
+    }
+
+    #[test]
+    fn truncate_handles_multibyte_boundary() {
+        // `é` is 2 bytes in UTF-8; build a string whose naïve cut at
+        // MAX_LEN would land mid-codepoint.
+        let mut big = "a".repeat(MAX_LEN - 1);
+        big.push('é');
+        big.push_str(&"b".repeat(64));
+        let s = scrub_for_log(&big);
+        // Must not panic; must produce valid UTF-8 (String guarantees).
+        assert!(s.contains("(truncated,"));
+    }
+
+    #[test]
+    fn idempotent_on_already_scrubbed_string() {
+        let once = scrub_for_log("Bearer abcdef api_key=hunter2");
+        let twice = scrub_for_log(&once);
+        assert_eq!(once, twice);
+    }
+}
diff --git a/src/openhuman/memory/tree/jobs/store.rs b/src/openhuman/memory/tree/jobs/store.rs
index d733dea59..c6641d7c7 100644
--- a/src/openhuman/memory/tree/jobs/store.rs
+++ b/src/openhuman/memory/tree/jobs/store.rs
@@ -22,6 +22,7 @@ use rusqlite::{params, Connection, OptionalExtension, Transaction};
 use uuid::Uuid;
 
 use crate::openhuman::config::Config;
+use crate::openhuman::memory::tree::jobs::redact::scrub_for_log;
 use crate::openhuman::memory::tree::jobs::types::{Job, JobKind, JobStatus, NewJob};
 use crate::openhuman::memory::tree::store::with_connection;
 
@@ -202,10 +203,14 @@ pub fn mark_failed(config: &Config, job: &Job, error: &str) -> Result<()> {
     with_connection(config, |conn| {
         let now_ms = Utc::now().timestamp_millis();
 
+        // `error` is free-form (anyhow chain or handler-supplied text) and
+        // may carry credential-shaped substrings; scrub before logging,
+        // but keep the original in the DB column for diagnostics.
+        let error_for_log = scrub_for_log(error);
         if attempts >= max_attempts {
             log::warn!(
                 "[memory_tree::jobs] terminal failure id={job_id} \
-                 attempts={attempts}/{max_attempts} err={error}"
+                 attempts={attempts}/{max_attempts} err={error_for_log}"
             );
             let n = conn.execute(
                 "UPDATE mem_tree_jobs
@@ -229,7 +234,7 @@ pub fn mark_failed(config: &Config, job: &Job, error: &str) -> Result<()> {
             let next_at = now_ms.saturating_add(backoff);
             log::info!(
                 "[memory_tree::jobs] retry id={job_id} attempt={attempts}/{max_attempts} \
-                 next_at_ms={next_at} err={error}"
+                 next_at_ms={next_at} err={error_for_log}"
             );
             let n = conn.execute(
                 "UPDATE mem_tree_jobs
@@ -531,6 +536,61 @@ mod tests {
         assert!(row.completed_at_ms.is_some());
     }
 
+    /// `mark_failed` scrubs only the log emission, not the persisted
+    /// `last_error` column. A reader of `mem_tree_jobs` should still see
+    /// the full anyhow / handler-supplied chain so they can root-cause
+    /// the failure; the scrub is a defense-in-depth for the log sink only.
+    #[test]
+    fn mark_failed_persists_full_error_unredacted() {
+        let (_tmp, cfg) = test_config();
+        let payload = AppendBufferPayload {
+            node: NodeRef::Leaf {
+                chunk_id: "c1".into(),
+            },
+            target: AppendTarget::Source {
+                source_id: "slack:#x".into(),
+            },
+        };
+        let mut nj = NewJob::append_buffer(&payload).unwrap();
+        nj.max_attempts = Some(1);
+        let id = enqueue(&cfg, &nj).unwrap().unwrap();
+
+        let claim = claim_next(&cfg, DEFAULT_LOCK_DURATION_MS).unwrap().unwrap();
+        let raw = "upstream returned 401: Bearer abc123.def-456 token rejected";
+        mark_failed(&cfg, &claim, raw).unwrap();
+
+        let row = get_job(&cfg, &id).unwrap().unwrap();
+        assert_eq!(row.status, JobStatus::Failed);
+        // The persisted column keeps the full original — the scrub is
+        // applied at log emission only.
+        assert_eq!(row.last_error.as_deref(), Some(raw));
+    }
+
+    /// Same contract for `mark_deferred`: the log line is scrubbed in
+    /// `worker::run_once`, but the persisted `last_error` keeps the
+    /// full handler-supplied reason for diagnostics.
+    #[test]
+    fn mark_deferred_persists_full_reason_unredacted() {
+        let (_tmp, cfg) = test_config();
+        let payload = AppendBufferPayload {
+            node: NodeRef::Leaf {
+                chunk_id: "c2".into(),
+            },
+            target: AppendTarget::Source {
+                source_id: "slack:#y".into(),
+            },
+        };
+        let nj = NewJob::append_buffer(&payload).unwrap();
+        let id = enqueue(&cfg, &nj).unwrap().unwrap();
+
+        let claim = claim_next(&cfg, DEFAULT_LOCK_DURATION_MS).unwrap().unwrap();
+        let raw = "rate_limited by api_key=hunter2; retry after 30s";
+        mark_deferred(&cfg, &claim, 0, raw).unwrap();
+
+        let row = get_job(&cfg, &id).unwrap().unwrap();
+        assert_eq!(row.last_error.as_deref(), Some(raw));
+    }
+
     #[test]
     fn recover_stale_locks_resets_running_rows() {
         let (_tmp, cfg) = test_config();
diff --git a/src/openhuman/memory/tree/jobs/worker.rs b/src/openhuman/memory/tree/jobs/worker.rs
index 79464f93b..da302ce4c 100644
--- a/src/openhuman/memory/tree/jobs/worker.rs
+++ b/src/openhuman/memory/tree/jobs/worker.rs
@@ -16,6 +16,7 @@ use tokio::sync::Notify;
 
 use crate::openhuman::config::Config;
 use crate::openhuman::memory::tree::jobs::handlers;
+use crate::openhuman::memory::tree::jobs::redact::scrub_for_log;
 use crate::openhuman::memory::tree::jobs::store::{
     claim_next, mark_deferred, mark_done, mark_failed, recover_stale_locks,
     DEFAULT_LOCK_DURATION_MS,
@@ -164,26 +165,31 @@ pub async fn run_once(config: &Config) -> Result<bool> {
             // claim toward the failure-attempt budget. `mark_deferred`
             // reverts the bump applied by `claim_next` so the row's
             // attempts counter stays where it was before this claim.
+            //
+            // `reason` is handler-supplied free-form text and may
+            // include upstream provider responses; scrub for log
+            // emission while keeping the original in DB state.
             log::info!(
                 "[memory_tree::jobs] deferred id={} kind={} until_ms={} reason={}",
                 job.id,
                 job.kind.as_str(),
                 until_ms,
-                reason
+                scrub_for_log(&reason)
             );
             mark_deferred(config, &job, until_ms, &reason)?;
         }
         Err(err) => {
             // Preserve the full anyhow cause chain in the persisted
             // last_error so a reader of mem_tree_jobs can see the root
-            // cause, not just the top-level message. Mirrors the {:#}
-            // log format used right above.
+            // cause, not just the top-level message. The log line gets
+            // the same chain after `scrub_for_log`, since anyhow chains
+            // commonly embed upstream HTTP bodies / auth headers.
             let message = format!("{err:#}");
             log::warn!(
-                "[memory_tree::jobs] job failed id={} kind={} err={:#}",
+                "[memory_tree::jobs] job failed id={} kind={} err={}",
                 job.id,
                 job.kind.as_str(),
-                err
+                scrub_for_log(&message)
             );
             mark_failed(config, &job, &message)?;
         }