Skip to content

feat(runtime): move to iii 0.26 observers - #1753

Open
rohitg00 wants to merge 56 commits into
mainfrom
feat/iii026-observers
Open

rohitg00 wants to merge 56 commits into
mainfrom
feat/iii026-observers

Conversation

@rohitg00

@rohitg00 rohitg00 commented Nov 11, 2026 •

Copy link
Copy Markdown
Owner

Stage 3 of the iii stack, on top of #1640. iii 0.26 removes triggers, so a worker that still binds with registerTrigger gets no events from a 0.26 engine. The engine, the state and http workers, and the SDK move together here.

Changes

  • iii-sdk, @iii-dev/helpers and the engine pin go to 0.26.0, with release SHA-256 for all seven platforms. The state and http registry workers go to 0.23.1 and 0.23.2 (the 0.26 observer releases) in the CLI, Docker and release gates.
  • src/runtime/bindings.ts observes http::on-request (route as args) and state::on-changed (scope and key filters) as soon as a route or subscription is registered. The registerTrigger fallback and the 30 s engine::functions::list polling are removed. Modes come from observer state, and boot no longer waits on them.
  • Observers use capture only. Failure audit observers use ["input", "error"] with an error filter and ["input", "output"] with an output.success == false filter.
  • Observer targets register through registerObserverTarget (src/runtime/observed.ts), which unwraps the Observed payload once. Handlers keep their typed bare event, and direct calls still work. HTTP handlers still receive the HttpRequest.
  • AGENTMEMORY_BINDINGS=trigger is treated as inline with a warning.
  • New docs/upgrades/iii-0.26.md. AGENTS.md, .env.example and the README upgrade links are updated.

Verification

  • build, skills:gen, docs:sync (no changes), skills:check, docs:check, eval typecheck
  • npm test: 4713 passed, 6 skipped. A full tsc -p tsconfig.json shows no new errors.
  • Live on an isolated HOME (instance 7): the CLI downloads and verifies engine 0.26.0, Compose starts state 0.23.1 and http 0.23.2, and REST is up in about 20 s on first run (6 s warm). Status reports 181 observers, 0 failed, with state, http and pipeline all on observe. Remember then smart-search works. A new memory reaches the viewer SSE through a state observer. Forget removes it from search. Stop takes 1.4 s, and an immediate restart (inside the 30 s observer grace window) re-binds every route with no OBSERVER_ID_CONFLICT and keeps the data.

Notes

  • The 0.26 engine logs four deprecation warnings for observers created by the upstream state and http workers themselves (configuration::on-updated, ade::ui::asset). They come from those workers' SDK, not from agentmemory.
  • Follow-up: the shared mockSdk test helper has no observe, so production bindings still tolerate a client without it. Giving the helper observe and making it required touches about a dozen test files, so it is left for a separate change.

Summary by CodeRabbit

  • New Features
    • Added dashboard warnings for duplicate workers and runtime health issues.
    • Added background compression and graph processing, with configurable runtime options.
    • Added improved tracking and cleanup for memory relationships and deleted data.
    • Added runtime status details for worker readiness, observer bindings, and recorded failures.
  • Bug Fixes
    • Improved recovery after state-worker interruptions and vector-index restoration.
    • CLI checks now identify outdated managed runtimes before upgrades.
    • Improved reporting when graph extraction completes only partially.
  • Documentation
    • Updated setup and deployment guidance for engine v0.26.0 and related worker versions, including automatic Windows installation instructions.
    • Updated the documented passing-test count to 5,000+ and expanded configuration guidance.

rohitg00 and others added 30 commits October 29, 2026 15:25
…nsport cap to 64 MiB

state 0.22.20 and http 0.22.7 are built on iii-sdk 0.24.6, which caps outgoing
messages at 16 MiB, so any whole-scope state read past that failed with
payload_too_large. state 0.22.22 and http 0.22.8 are iii-sdk 0.25.0 builds that
fragment messages up to 64 MiB, so the payload guard moves to 60 MiB.

Graph index migration now appends the underlying error to its storage-failure
message, AGENTMEMORY_LOCAL_DISPATCH=false sends internal calls back through the
engine for full traces, and gate:large-store measures large scopes against a
real Compose runtime. The startup gate gains --repeat N for cold-boot loops.

Generated with [Devin](https://devin.ai)

Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Signed-off-by: Rohit Ghumare <ghumare64@gmail.com>
…errupts in repeat mode

The large-store gate now waits for the agentmemory worker to appear in
engine::workers::list before taking the baseline, and treats an empty or
failed observation as unstable instead of a pass. Repeat mode in the startup
gate forwards SIGINT/SIGTERM/SIGHUP to the active child's process group so
its teardown runs, escalating to SIGKILL on a second signal. The Docker
Compose environments forward AGENTMEMORY_LOCAL_DISPATCH, and the
registration-defaults tests no longer depend on the caller's environment.

Generated with [Devin](https://devin.ai)

Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Signed-off-by: Rohit Ghumare <ghumare64@gmail.com>
Even with 64 MiB transport messages, a single state::list of the insights or
graph scopes can exceed the limit (524 MB insights, 73 MB graph nodes in the
large-store gate). StateKV gains listKeys, getMany, iterate and listPage built
on state::list_keys plus batched state::get, and switches to state::get_many
when the state worker provides it. Every reader of an unbounded scope now goes
through them, and a guard test blocks new whole-scope reads.

Mesh export pages with a stable id cursor and the sync client follows it to
the end, stamping lastSyncAt with the sync start time only after the final
page. Export no longer swallows collection read failures, which silently
dropped semantic and insight data when the read failed. Graph index migration
discovers ids through listKeys.

Generated with [Devin](https://devin.ai)

Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Signed-off-by: Rohit Ghumare <ghumare64@gmail.com>
… imports consistent

Mesh export now emits a cursor when a page fills exactly at a collection
boundary, so later collections are no longer dropped. The sync watermark is
the per-peer start time and only advances when every requested direction
succeeded, so a failed push is retried. The speculative state::get_many path
is removed until the state worker offers a byte-bounded batch protocol.
collectAll restores global insertion order on the redis backend. Replace
imports read every row they will delete before deleting anything, so a read
failure no longer leaves a half-deleted store.

Generated with [Devin](https://devin.ai)

Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Signed-off-by: Rohit Ghumare <ghumare64@gmail.com>
…vers

All 152 HTTP routes and 12 state reactions now go through one bindings layer.
With state 0.22.22 and http 0.22.8 it observes state::on-changed and
http::on-request, so agentmemory no longer uses the deprecated registerTrigger.
It falls back to registerTrigger per provider when the observable function is
missing, never binds a spec both ways, and AGENTMEMORY_BINDINGS=trigger keeps
the old path. State filters cover graph generation scopes and key-only
handlers.

Boot waits for the state worker with a bounded, non-fatal probe
(AGENTMEMORY_STATE_READY_TIMEOUT_MS). Index and capture recovery run again when
the state worker comes back, detected by state observers returning to active or
an engine::on-workers-changed observer, with a probe tick in trigger mode.
Duplicate agentmemory workers are reported in status, the viewer and doctor.

Generated with [Devin](https://devin.ai)

Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Signed-off-by: Rohit Ghumare <ghumare64@gmail.com>
Owner-delivered http::on-request observers default to 64 deliveries in flight,
so the 65th concurrent request failed with OVERFLOW where trigger-bound routes
had queued. Route observers now set max_in_flight from
AGENTMEMORY_ROUTE_MAX_IN_FLIGHT (default 1024); a 300-request burst returns 201
for every request. Also drops a stray code fence in AGENTS.md.

Generated with [Devin](https://devin.ai)

Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Signed-off-by: Rohit Ghumare <ghumare64@gmail.com>
When boot runs before the state worker answers, vector recovery was paused
until restart. The re-arm now reloads the checkpoint after the store comes
back, but only when the boot load failed because the checkpoint was
unavailable. The trigger-mode probe starts from the boot readiness result. A
re-arm that arrives inside the throttle window is retried once the window ends
instead of being dropped. The viewer's duplicate-worker alert reads the
status report it is actually given.

Generated with [Devin](https://devin.ai)

Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Signed-off-by: Rohit Ghumare <ghumare64@gmail.com>
With iii 0.25 observers available, background work no longer goes out as
unbounded fire-and-forget calls. Compression runs from an observer on
mem::observe capped at AGENTMEMORY_COMPRESS_MAX_CONCURRENT (default 2), and a
periodic mem::compression-sweep re-dispatches observations still pending or
failed. Session-stop graph extraction, reflect and consolidation become
observers on event::session::stopped. Graph extraction only sends observations
newer than a per-session marker (KV.graphExtractMarks), with an optional
cooldown. AGENTMEMORY_GRAPH_STRUCTURAL=false turns off the heuristic graph
pass. Saved memories reach graph extraction through an observer on
mem::remember. Without observers, the existing inline paths are unchanged.

Generated with [Devin](https://devin.ai)

Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Signed-off-by: Rohit Ghumare <ghumare64@gmail.com>
This branch carried an earlier copy of the paged-reads work that main now
holds as #1636. Keep main for those files and apply only the observer
binding commits on top. Index recovery keeps the single-flight wrapper
and main's superseded-revision handling, and the boot wait from main
runs before the observer status probe.
- Retry an unavailable vector checkpoint before index recovery on a state
  worker return, and rebuild once it is restored.
- Merge a retried checkpoint into the live vector index instead of
  replacing it, so vectors embedded after boot stay searchable.
- Count trigger-mode routes, bound each function-list probe by the
  detection deadline, and report rejected observers in status.
- Load the status report from the dashboard so the duplicate-worker
  alert shows without visiting Health.
- Drop the ?collections hint the mesh export never honored, detect a
  truncated success:false body in the large-store gate, and leave
  CHANGELOG to the release PR.
…pipeline

CHANGELOG follows the base branch; release PRs carry the entry.
…nd late compression

- Leave the post-session graph marker unchanged when graph extraction
  reports success false.
- Tie the marker to the graph generation so a graph reset re-extracts
  sessions extracted before it.
- Use compressedAt when present, so a stop before LLM compression does
  not hide the compressed concepts and files from a later extraction.
- Pass a saved memory's files to graph extraction.
…audit

CHANGELOG follows the base branch; release PRs carry the entry.
…radiction floor

- Delete facets for a flushed batch with one scan per target type
  instead of one full facet scan per deleted record.
- Requeue a delete whose target read failed, so its facets are cleaned
  up on the next flush instead of being dropped.
- Track each contradiction's raw penalty against a base confidence and
  apply the change in the floored value, so undoing one contradiction
  does not lift confidence while another still holds it at the floor.
- Cover search and vector restoration on the final supersedes undo.
… current

- Re-run index recovery on a state worker return only after an incomplete
  keyword rebuild, a paused checkpoint or a failed recovery run. A backfill
  paused at its per-start cap is no longer resumed without opt-in.
- Rebuild a stale session index on a state worker return.
- Keep runtime.stateWorker.ready in step with probe and observer state.
- Refresh the duplicate-worker report on the 30 s tick, and read the
  engine worker id field.
- Send the doctor status request with headers for the daemon base URL.
… finite

- Re-add an un-superseded memory to the vector index with its session,
  embed text and memory context; the previous two-argument call never
  embedded it.
- Treat a memory without a stored confidence as 1 in the contradicts
  reaction, so the stored value never becomes NaN, and type the field.
- Rebuild the keyword index after a replace import from the store passed
  to the handler; the previous closure referenced an undefined name.
…leted

- Re-run index recovery on a state worker return when the vector
  checkpoint has a save error.
- Remember vectors deleted while absent from the live index, so a
  retried checkpoint merge does not restore them.
- Bind the lesson-get route through Bindings.
- Give the doctor status request 8 s, above the server check budget.
A contradiction on a memory with no stored confidence started from 1, so
the stored value could rise above the type salience and promote a
contradicted fact in retention. Start from 0.5, the default salience.
- Snapshot the live vector index before a retried checkpoint load and
  re-apply it after pending-log replay, so replay cannot overwrite newer
  embeddings or restore deleted vectors.
- Track removed ids only while a boot checkpoint retry is pending.
- Re-arm recovery after a backfill batch that embedded nothing and had
  failures, which stays within the per-start cap.
coderabbitai[bot]

This comment was marked as resolved.

The direct pipeline calls were turned off as soon as the SDK supported
observe(), even when a pipeline observer stayed pending or was rejected.
Compression, the post-session graph, reflection and consolidation then
stopped with no fallback. Each step now runs as a direct call unless its
own observer is active, so a partly attached pipeline neither drops work
nor runs a step twice. runtime.bindings.reasons lists the steps that use
direct calls.

The compression sweep sent failed rows in key order with no limit, so
permanent failures were retried every run and pushed newer pending rows
out of the batch. It now sends pending rows first, retries a failed row
at most 5 times and skips rows whose source is gone.

A route or state observer rejected with OBSERVER_ID_CONFLICT, which
happens while the engine still holds the ids of an earlier connection,
is now observed again after 10, 20, 40 and 80 seconds before it counts
as rejected. Before, all routes stayed at 404 until a manual restart.

The observer-rejected status message and the iii 0.26 upgrade notes
describe the fallback, and the notes no longer say that the CLI status
command shows pending providers.

Signed-off-by: Rohit Ghumare <ghumare64@gmail.com>
…ile it runs

The Compose state recorded the worker versions it launched but never read
them back, and reconcile checked only the iii binary on disk. After
`agentmemory upgrade` replaced ~/.agentmemory/bin/iii with 0.26.0, a still
running 0.25.1 engine with 0.22.x workers passed the check and the CLI
reported ready while every worker route returned 404.

- Record the launched engine version (engineVersion) in engine-state.json
  next to workerVersions.
- In reconcilePersistedComposeEngine, compare the persisted engine and
  worker versions with the current pins and exit with an agentmemory stop
  instruction on a mismatch. State files from older CLIs without
  engineVersion still fail on a workerVersions mismatch.
- Refuse `agentmemory upgrade` and the doctor installer fixes while a
  managed Compose engine is running, and tell the user to run
  agentmemory stop first.

Signed-off-by: Rohit Ghumare <ghumare64@gmail.com>
devin-ai-integration[bot]

This comment was marked as resolved.

… unavailable boot checkpoint

resumeAfterUnavailable no longer sets recoveryReady. Checkpoint saves and
the pending log stay paused until prepareBackfill finishes on the reloaded
checkpoint. Before, a live vector write during a slow reload could save
checkpoint metadata before loadBuckets restored the bucket assignments.

resumeAfterUnavailable now clears the vector leg error that pauseRecovery
set. Before, a restored checkpoint that needed no vector write kept the
stale error, and completeBackfill reported recovery as paused.

onStateWorkerReturn takes an optional pending check. probeTick runs the
re-arm callback while that check is true, also when the state worker never
left. The boot code passes a check for an unavailable boot checkpoint, so a
temporary checkpoint read failure is retried every 30 s instead of staying
paused until the next worker event.

Signed-off-by: Rohit Ghumare <ghumare64@gmail.com>
…bservers in inline mode

mem::graph-extract now returns partial: true and the LLM error when the
structural pass persisted nodes but the LLM pass failed.
mem::post-session-graph treats that result as a failure, so the per-session
marker does not advance and the observations are extracted again on the
next session stop.

When the post-session graph observer is not active, event::session::stopped
now calls mem::post-session-graph instead of mem::graph-extract. The direct
path then uses the same marker, cooldown and fresh-observation selection as
the observer path, so observations extracted by the fallback are not
extracted again after the observer becomes active.

observeStateChanges now registers its state::on-changed observer in inline
mode too. AGENTMEMORY_BINDINGS=inline only moves the post-capture pipeline
inline; the delete cascade and relation reaction observers must stay active
so deleted rows leave the search index and relation effects still apply.

Signed-off-by: Rohit Ghumare <ghumare64@gmail.com>
coderabbitai[bot]

This comment was marked as resolved.

…retries and cycles

Worker shutdown now awaits flushDeleteCascade (bounded by the shutdown
flush timeout) before the SDK shuts down. Queued facet cleanup and
observationCount reconciliation no longer get lost when the process exits
inside the 500 ms batch window.

The supersedes cycle guard now walks the existing supersedes reaction
marks from the new source. It rejects the relation when the target
already supersedes the source through any chain, not only through a
direct reverse relation. Before, A->B->C->A set isLatest to false on all
three memories.

Relation reactions now write the mark first with a pending snapshot of
the target fields, then the memory, then the final mark. When a delivery
finds a pending mark, it completes the write before it continues. A
failed mark write after the memory write therefore cannot apply a second
strength reduction or record already-changed values as the baseline. The
undo path uses the same commit.

mem::on-record-deleted reads the stored value before it removes BM25 and
vector entries. It skips the removal when the key exists again and is not
marked deleted, so a delayed delete after a replace import does not drop a
restored record from search. Deleted lessons are still removed.

Signed-off-by: Rohit Ghumare <ghumare64@gmail.com>
… and Windows docs

- agentmemory upgrade and the doctor installer fixes checked only the
  selected instance's engine-state.json. Instances above 0 keep that file
  under <data dir>/instance-N, so a running --instance 1 did not block an
  upgrade that replaces the shared ~/.agentmemory/bin/iii. The guard now
  reads the Compose state of the default runtime dir and of every
  instance-1..50 data dir (resolved with the same --data-dir and env
  rules), and reports a running engine when the Compose PID is alive or
  the recorded REST port answers.
- compressionSweepMs() returned 0 for an empty
  AGENTMEMORY_COMPRESSION_SWEEP_MS, which the Compose files pass when the
  variable is unset, so the recovery sweep never ran. An empty or blank
  value now means the 10-minute default; an explicit 0 still disables it.
- docs/upgrades/iii-0.25.md said the bindings fall back to registerTrigger
  and that trigger mode keeps the inline fan-out. On iii 0.26 there is no
  fallback and AGENTMEMORY_BINDINGS=trigger is treated as inline. The text
  now says so and links the iii 0.26 upgrade guide.
- The translated READMEs said native Windows needs a manual engine
  install and that the CLI does not extract the Windows ZIP. The CLI
  downloads, verifies and extracts iii.exe into %USERPROFILE%\.agentmemory\bin.
  All 32 translations now match README.md, including the vi-VN install
  requirements line. docs:sync refreshed the generated counts.

Signed-off-by: Rohit Ghumare <ghumare64@gmail.com>
devin-ai-integration[bot]

This comment was marked as resolved.

…andle

activate() subscribed only to the handles that existed when it started.
When an OBSERVER_ID_CONFLICT retry replaced a route or state handle, it
kept waiting on the old one and resolved only at the 30 second timeout.
New handles are now passed to active waiters, which also check at once
in case the replacement is already active.

Signed-off-by: Rohit Ghumare <ghumare64@gmail.com>
coderabbitai[bot]

This comment was marked as resolved.

…unction count

mem::graph-extract returns success with partial: true when the LLM pass
fails after the structural pass stored nodes. The failure observers
matched only thrown errors and success: false, so these LLM failures
never reached mem::record-failure. A third failure observer now matches
output.partial on mem::graph-extract, and record-failure maps its id to
graph-extract-failed.

AGENTS.md listed "322+ iii functions", a form docs:sync does not rewrite.
It now uses the exact synced count like the other stats lines.

Signed-off-by: Rohit Ghumare <ghumare64@gmail.com>
@rohitg00

Copy link
Copy Markdown
Owner Author

Wanna take a look on this? @guibeira

Signed-off-by: Rohit Ghumare <ghumare64@gmail.com>
…moval and state recovery

- The upgrade guard resolved sibling instances from AGENTMEMORY_DATA_DIR
  after the CLI had already set it to the selected instance, so
  `upgrade --instance 3` looked in instance-3/instance-1. It now uses the
  value from before that override.
- When the restoration check for a deleted key failed to read, the record
  was treated as deleted and removed from both indexes, even if it was
  live. The index entries now stay, and the delete flush removes them
  once it confirms the record is gone.
- State recovery subscribed only to the state handles that existed when
  it started. A handle that replaced a conflicting one never triggered
  the re-arm. Bindings.onStateHandle now passes every new state handle to
  recovery, which also handles a replacement that is already active.

Signed-off-by: Rohit Ghumare <ghumare64@gmail.com>
@guibeira

Copy link
Copy Markdown
Contributor

Wanna take a look on this? @guibeira

Sure!

Signed-off-by: Rohit Ghumare <ghumare64@gmail.com>
… back off delete flush retries

Edge reactions and mem::relate read a memory under its mem:<id> lock and
wrote it back later. Memory deletes hold mem:remember, so a delete landing
between that read and write let the writer put the memory back, and the
delete cascade then treated it as restored and kept its index entries.
Both writers now take mem:remember inside mem:<id> for the read-modify-write.

The delete cascade rescheduled a failed flush every 500 ms forever while
the state worker was unreachable. Retries now double from 500 ms up to
30 s and reset after a flush that succeeds.

Signed-off-by: Rohit Ghumare <ghumare64@gmail.com>

@devin-ai-integration devin-ai-integration Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Devin Review found 3 potential issues.

Devin Review

Comment on lines +35 to +37
const since = marker ? Date.parse(marker.lastObservationTimestamp) : 0;
const observations = await collectAll<CompressedObservation>(kv, KV.observations(data.sessionId));
const fresh = observations.filter((o) => o.title && Date.parse(graphStamp(o)) > since);

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🟡 Imported observations excluded from graph extraction

After a session has an extraction marker, fresh excludes imported observations dated before that marker. Later session stops never extract those observations, even when newer observations trigger another graph pass.

Learn more

The extraction marker is a timestamp watermark, not a record of processed observations. mem::import can merge observations into an existing session while preserving their original timestamps and compression dates. Such observations can be new to this graph while older than its watermark. Subsequent extraction keeps excluding them, whereas the previous session-stop implementation extracted all titled observations.

Example: A session's graph marker is November 15. Merge an observation from November 10 into that session, then capture a November 16 observation and stop the session. Only the November 16 observation reaches graph extraction; the imported observation never does.

Recommended fix: Track extracted observation revisions, or invalidate the affected session's extraction marker when importing observations. Cover merge imports into sessions with an existing marker and a later successful session stop.

Devin Review


Was this helpful? React with 👍 or 👎 to provide feedback.

Comment on lines +47 to +52
if (isGraphExtractionEnabled()) {
bindings.observePipeline(
{ function_id: "mem::remember", capture: ["input", "output"] },
"mem::post-remember-graph",
{ id: "agentmemory:pipeline:post-remember-graph", max_in_flight: 1 },
);

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🟡 Saved-memory graph processing missing inline fallback

With graph extraction enabled, observePipeline skips saved-memory graph processing in inline mode. mem::remember has no direct fallback, so saved memories never reach graph extraction through this flow.

Learn more

observePipeline returns without registering anything in inline mode. Unlike compression and post-session processing, mem::remember never checks whether its graph observer is active or directly invokes the replacement step. Pending or rejected observer registration also leaves this stage without a fallback.

Example: Set AGENTMEMORY_GRAPH_EXTRACTION=true and AGENTMEMORY_BINDINGS=inline, then save a memory containing file and concept links. The save succeeds, but no graph extraction runs for that memory.

Recommended fix: Add a direct fallback after a successful save when observing("mem::post-remember-graph") is false. Share the observation-building logic between observer and direct paths, and test inline and inactive-observer cases.

Devin Review


Was this helpful? React with 👍 or 👎 to provide feedback.

Comment on lines +74 to +83
const content = typeof data.input?.content === "string" ? data.input.content : "";
const observation = {
id: memory.id,
title: memory.title || content.slice(0, 120) || memory.type,
narrative: content,
type: "compressed_observation",
timestamp: memory.createdAt,
sessionId: `memory:${memory.id}`,
hookType: "remember",
concepts: data.input?.concepts ?? memory.concepts ?? [],

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🟥 Private memory content sent to graph provider

With LLM graph extraction enabled, mem::post-remember-graph forwards captured input containing secrets or private blocks. withWriteScrubbing scrubs only the save handler's copy, not captured invocation input. The graph provider receives content excluded from the saved memory.

Devin Review


Was this helpful? React with 👍 or 👎 to provide feedback.

This branch was successfully deployed

1 active deployment
Preview — 87cd6d84 Deployed Nov 16, 2026 by vercel[bot]
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants