Operations

Memory operations keep the retrieval corpus small enough to search and explicit enough to repair. The implementation lives in operations.ts, jobs.ts, items.ts, and runtime/src/temporal/memory.

Job types

runMemoryMaintenanceJob() supports five job types.

Type	Required scope	Behavior
`consolidate`	`agentWallets` or `agentWallet`	Marks duplicate content/source rows as `metadata.status: "superseded"` and lowers their `decayScore` to at most `0.05`.
`patterns_extract`	`agentWallet`	Reads transcripts in a time window and extracts repeated tool-call sequences.
`archive_create`	`agentWallet`	Creates a compressed archive from vector rows in a date range.
`decay_update`	None	Recomputes vector `decayScore` with the requested half-life.
`cleanup`	None	Deletes stale low-decay vectors, expired sessions, and old transcripts.

Jobs are stored in memory_jobs with running, completed, or failed status.

Inline and Temporal execution

The low-level jobs.ts runner executes maintenance inline. The Temporal service runs the same jobs through Temporal when execution: "temporal" is requested and Temporal is configured.

Execution	Runtime behavior
`inline`	Runs in the runtime process and writes one `memory_jobs` record.
`temporal`	Starts a Temporal workflow, stores workflow ids on the job record, and refreshes state from Temporal queries.

Temporal memory workflows send heartbeats, retry activities, support pause/resume signals, and continue-as-new during long consolidation batches.

Schedules

Memory schedules are managed through /api/memory/schedules.

Schedule	Frequency	Workflow
Daily consolidation	`MEMORY_DAILY_CONSOLIDATION_CRON`	`memoryConsolidationWorkflow`
Daily pattern extraction per agent	`MEMORY_DAILY_CONSOLIDATION_CRON`	`patternExtractionWorkflow`
Weekly archive per agent	`MEMORY_WEEKLY_ARCHIVE_CRON`	`archiveCreationWorkflow`
Hourly decay	`MEMORY_HOURLY_DECAY_CRON`	`decayUpdateWorkflow`
Daily cleanup	`MEMORY_DAILY_CONSOLIDATION_CRON`	`memoryCleanupWorkflow`

The default overlap policy is SKIP, except weekly per-agent archives use BUFFER_ONE.

Pattern extraction

extractExecutionPatterns() reads session_transcripts, extracts tool-call names, groups identical sequences, and stores frequent sequences as tool_sequence patterns.

{
  "patternType": "tool_sequence",
  "trigger": {
    "type": "context",
    "value": "search_call"
  },
  "steps": [
    { "action": "search_call", "order": 0 },
    { "action": "models_call", "order": 1 }
  ],
  "successRate": 0.65,
  "executionCount": 3
}

The runtime can validate a pattern and promote it into a skills document. Promotion creates a learned skill with a pattern trigger, the extracted tool list, and success metadata.

Field	Meaning
`archiveId`	Runtime-generated archive id.
`agentWallet`	Agent scope.
`dateRange`	Source vector range.
`content`	JSON payload or gzip base64 payload.
`compressed`	Whether gzip compression was used.
`metadata.entryCount`	Number of vector rows included.
`ipfsCid`	Optional Pinata IPFS CID after sync.

Item lifecycle

The item APIs work against memory collection rows.

Operation	Behavior
`getMemoryItem()`	Reads an active row by `vectorId` and optional scope.
`updateMemoryItem()`	Updates content, metadata, retention, confidence, or status. Re-embeds when content changes.
`deleteMemoryItem()`	Soft-deletes by setting `metadata.status: "deleted"` and `decayScore: 0`; `hardDelete` removes the row.
`resolveMemoryConflict()`	Records conflict metadata and can mark a row `superseded`.

Soft-deleted and superseded rows are excluded by active-layer filters. Hard deletes remove the row permanently.

Cleanup

cleanupExpiredMemories() deletes:

Data	Condition
Vectors	`createdAt` older than the cutoff, `decayScore < 0.2`, and `metadata.retention != "pinned"`.
Sessions	`expiresAt` in the past or `lastAccessedAt` older than the cutoff.
Transcripts	`createdAt` older than the cutoff.

After cleanup, the runtime invalidates memory caches globally.

Endpoint map

Family	Routes
Loop	`POST /api/memory/context/assemble`, `POST /api/memory/turns/record`, `POST /api/memory/remember`, `POST /api/memory/loop`
Manifests	`GET /api/memory/workflows`, `GET /api/memory/workflows/{workflowId}`
Patterns and skills	`GET /api/memory/patterns`, `GET /api/memory/patterns/{patternId}`, `POST /api/memory/patterns/{patternId}/validate`, `POST /api/memory/patterns/{patternId}/promote`, `GET /api/memory/skills`, `GET /api/memory/skills/{skillId}`
Sessions and archives	`POST /api/memory/transcripts/index`, `GET/PATCH /api/memory/sessions/{sessionId}/working`, `POST /api/memory/sessions/{sessionId}/compress`, `POST /api/memory/archives/{archiveId}/sync`
Schedules	`GET/POST/DELETE /api/memory/schedules`, `POST /api/memory/schedules/{scheduleId}/pause`, `resume`, `trigger`
Items	`POST /api/memory/items/search`, `GET/PATCH/DELETE /api/memory/items/{id}`, `POST /api/memory/conflicts/{id}/resolve`
Jobs and evals	`POST /api/memory/jobs`, `GET /api/memory/jobs/{jobId}`, `POST /api/memory/evals/runs`
Compatibility	`POST /api/memory/add`, `POST /api/memory/search`, `GET /api/memory/{agentWallet}`, `POST /api/memory/vector-search`, `POST /api/memory/vector-index`, `POST /api/memory/transcript-store`, `GET /api/memory/transcript-get/{id}`, `POST /api/memory/rerank`, `POST /api/memory/layers/search`

Overview

Memory

Harness

Tools

Job types

Inline and Temporal execution

Schedules

Pattern extraction

Archives

Item lifecycle

Cleanup

Endpoint map

​Job types

​Inline and Temporal execution

​Schedules

​Pattern extraction

​Archives

​Item lifecycle

​Cleanup

​Endpoint map

​Related

Job types

Inline and Temporal execution

Schedules

Pattern extraction

Archives

Item lifecycle

Cleanup

Endpoint map

Related