Automations

Automations are Memory Layer's built-in loop workflows. They are registered by the local service, scoped by project or repository, and recorded as auditable runs with traces, approvals, and proposed outputs.

They are not external bots and they are not repo-local agent skills. Skills tell a coding agent how to use Memory Layer; automations are service-owned workflows that can be run manually or routed from trusted events.

How they work

memory loops list
memory loops show memory_hygiene --project memory
memory loops run memory_hygiene --project memory --dry-run --reason "inspect cleanup candidates"
memory loops runs --project memory
memory loops inspect <run-id>

Each automation has:

Field	Meaning
Loop ID	Stable identifier used by CLI, API, and run records.
Default mode	Starting behavior before project-specific settings override it.
Risk level	Operator signal for how much review the loop deserves.
Supported triggers	Events that can route to the loop, such as `manual`, `schedule`, `memory_changed`, or `ci_failed`.
Capabilities	What evidence the loop is allowed to read or request.
Outputs	Reports, proposals, task packs, draft metadata, or metrics produced by the run.

Loop outputs still use Memory Layer APIs and approval flows. A loop does not get a private write path to memories or code.

Built-in loops

Loop	Default	Risk	Triggers	Outputs
`context_pack_refresh`	`suggest_only`	Low	`manual`, `repo_docs_changed`, `memory_changed`	Context-pack diff and memory proposals.
`memory_hygiene`	`suggest_only`	Medium	`manual`, `schedule`, `memory_changed`	Merge, deprecate, and link proposals plus a hygiene report.
`ci_failure_triage`	`observe`	Medium	`manual`, `ci_failed`	CI triage report and optional follow-up task proposal.
`agent_ready_issue_triage`	`suggest_only`	Low	`manual`, `issue_labeled`, `issue_created`	Issue report and task-pack proposal.
`draft_pr`	`draft_output`	High	`manual`, `agent_ready_issue`	Draft PR metadata, checks, and memory proposals.
`reviewer_drift_detection`	`suggest_only`	Medium	`manual`, `pull_request_opened`, `pull_request_updated`	Review report and architecture memory proposal when appropriate.
`skill_mining`	`suggest_only`	Medium	`manual`, `run_succeeded`, `pull_request_merged`	Learned-skill memory proposal.
`memory_eval`	`observe`	Low	`manual`, `schedule`, `retriever_changed`	Retrieval and context-quality metrics.

Loop details

`context_pack_refresh`

Builds or refreshes the context pack a future run would use for a project or repository. When it is not blocked by policy, it records the pack and can propose architecture summaries, command lists, conventions, module maps, and stale-memory warnings. Proposals stay pending until reviewed.

`memory_hygiene`

Inspects active memory summaries and metadata for likely duplicates, low-confidence memories, stale low-importance memories, and related same-tag memories. It emits proposals using summaries and IDs only, so full canonical text is not copied into hygiene reports.

`ci_failure_triage`

Classifies attached CI evidence as likely regression, flaky, environmental, dependency-related, or unknown. In observe mode it records a report only. In suggest-only mode it can propose a follow-up task for suitable regression or dependency failures. It does not write code.

`agent_ready_issue_triage`

Reads an issue payload, classifies ambiguity and implementation risk, suggests labels such as agent-ready, needs-human-clarification, or needs-design, estimates likely files and tests, and creates a task-pack proposal for suitable low-risk work.

`draft_pr`

Creates isolated draft implementation work for approved low-risk issue payloads. Sensitive areas such as auth, billing, security, migrations, secrets, deletion, or infrastructure stop at approval. Approved runs use generated worktrees and branches and never push directly to main.

`reviewer_drift_detection`

Reviews changed files and diff metadata against remembered architecture, conventions, and safety constraints. It flags unrelated changes, missing tests, hidden behavior changes, security-sensitive diffs, and architecture drift. It does not modify code.

`skill_mining`

Reads successful run or accepted PR payloads and looks for reusable development recipes. Suitable payloads become pending learned-skill memory proposals with applicability conditions, recipe steps, commands, validation evidence, and source links.

`memory_eval`

Produces an internal run report for retrieval and context quality. It tracks precision/recall proxies, stale-memory injection rate, contradiction rate, accepted proposal rate, useful-run rate, cost per useful run, and context-pack comparison data.

Controls and safety

Use project-scoped settings unless you intentionally want broader behavior:

memory loops enable memory_hygiene --project memory --mode suggest_only --explicit-user-approval
memory loops pause memory_hygiene --project memory --until 2026-06-16T09:00:00Z
memory loops disable memory_hygiene --project memory --reason "pause while reviewing output quality"

The global kill switch blocks routed non-manual runs. Manual memory loops run ... calls still record a trigger event, but they are deliberate operator actions and are not blocked by the kill switch.

Approval-gated actions include memory mutation, repository writes, and runner invocation. Forbidden actions include direct pushes to main, deploys, secret access, destructive migrations, and enabling loops from inside a loop.

Where to inspect them

Surface	Use it for
Browser UI Automations tab	Inspect definitions, settings, recent runs, approvals, context packs, and memory proposals.
`memory loops list`	Show registered loop definitions.
`memory loops run`	Start a manual dry-run or run.
`memory loops approvals`	Review pending approval requests.
`memory loops memory-proposals`	Review proposals emitted by loop runs.
`memory loops inspect <run-id>`	Inspect a specific run, traces, and outputs.

Read Browser UI, CLI reference, or Operations.