> ## Documentation Index > Fetch the complete documentation index at: https://docs.memloom.dev/llms.txt > Use this file to discover all available pages before exploring further. # Run the consolidation pass > Reviews the store, repairs what SQL proves wrong, folds duplicate entity names, and asks about the rest. Every run is one undoable unit: the report names the run id, and `/memory/reconcile/{id}/revert` puts back exactly what it did. `mode` defaults to `dry_run`, which changes nothing and never calls a model, returning what the paid passes would have cost in `estimate`. **The `memloom reconcile` CLI command defaults the other way and applies.** Passes come from the saved settings unless `passes` overrides them: two free ones on, three paid ones off. With the contradiction re-check on a run takes minutes, so use the stream variant rather than holding this request open. See [Reconciliation](https://docs.memloom.dev/concepts/reconciliation). ## OpenAPI ````yaml /openapi.yaml post /memory/reconcile openapi: 3.1.0 info: title: memloom local API version: 0.1.0 description: > The HTTP API served by the `memloom serve` daemon on your machine. The CLI, the MCP server, and the viewer all route through this API. The daemon is the single owner of the store, so every client shares one consistent view of your memories. No authentication: the daemon binds to `127.0.0.1` only and is reachable solely from your own machine. Browser clients on other localhost ports are allowed via CORS. servers: - url: http://127.0.0.1:4319 description: Local memloom daemon security: [] paths: /memory/reconcile: post: summary: Run the consolidation pass description: > Reviews the store, repairs what SQL proves wrong, folds duplicate entity names, and asks about the rest. Every run is one undoable unit: the report names the run id, and `/memory/reconcile/{id}/revert` puts back exactly what it did. `mode` defaults to `dry_run`, which changes nothing and never calls a model, returning what the paid passes would have cost in `estimate`. **The `memloom reconcile` CLI command defaults the other way and applies.** Passes come from the saved settings unless `passes` overrides them: two free ones on, three paid ones off. With the contradiction re-check on a run takes minutes, so use the stream variant rather than holding this request open. See [Reconciliation](https://docs.memloom.dev/concepts/reconciliation). operationId: reconcile requestBody: required: false content: application/json: schema: $ref: '#/components/schemas/ReconcileInput' responses: '200': description: What the run found and what it changed. content: application/json: schema: $ref: '#/components/schemas/ReconcileReport' '400': $ref: '#/components/responses/ValidationError' '403': description: > `RECONCILE_ENABLED=0` is set on the daemon and `mode` is `apply`. The kill switch refuses any run that was asked to act; a dry run still works. content: application/json: schema: type: object properties: error: type: string example: >- reconciliation is set to report only (RECONCILE_ENABLED=0). Remove it from your memloom config and restart the daemon to let a run act. '409': description: >- Another applying run is already going. Watch it or wait for it to finish. content: application/json: schema: type: object properties: error: type: string example: >- memloom: a reconcile run is already going. Watch it in the Console, or wait for it to finish. components: schemas: ReconcileInput: type: object description: What to run. Every field is optional. properties: mode: type: string enum: - dry_run - apply default: dry_run description: >- `dry_run` reports and changes nothing, and never calls a model. `apply` acts. trigger: type: string enum: - manual - idle - startup default: manual description: Recorded on the run, so the history says who started it. passes: type: array description: >- Run exactly these passes, ignoring the saved settings. Omit to use the settings, which is what every shipped surface does. This overrides the settings for the three paid passes too, so a caller can start a run that spends money while the Settings toggles are off. `RECONCILE_ENABLED=0` still refuses it. items: type: string enum: - invariants - entities - llm_entities - llm_conflicts - llm_recheck budgetUsd: type: number exclusiveMinimum: 0 maximum: 50 description: >- Keep re-checking past the per-run ceiling until nothing is due or this much has been billed, measured against the provider's own reported cost. Omit for one page and stop, which is the default. Capped at 50 because a typo here spends real money. ReconcileReport: type: object description: >- One run's own words. This is what the CLI prints and the Settings tab renders. properties: run: $ref: '#/components/schemas/ReconcileRun' actions: type: array items: $ref: '#/components/schemas/ReconcileAction' estimate: $ref: '#/components/schemas/ReconcileEstimate' passes: type: array items: type: string description: The passes this run actually ran. heldBack: type: object description: >- Findings recorded but not surfaced, because a per-run cap was reached. properties: retire: type: integer question: type: integer conflict: type: integer entities: $ref: '#/components/schemas/EntityResolutionResult' arbitration: type: object description: Present when the uncertain entity pairs pass ran. properties: calls: type: integer folded: type: integer rejected: type: integer description: >- Pairs the model said are different things. Recorded, so it is never asked again. unsure: type: integer description: Left pending for a human. autoResolved: type: object description: Present when the memory conflicts pass ran. properties: examined: type: integer resolved: type: integer recheck: type: object description: Present when the contradiction re-check ran. properties: window: type: integer description: Beliefs this run swept. calls: type: integer claimed: type: integer description: >- Contradictions the model claimed, before the quotes were checked. verified: type: integer description: >- Findings whose quotes were found in both memories, and were recorded. remaining: type: integer description: Beliefs still due when the run stopped. spentUsd: type: number description: Billed for this pass, from the provider's own figures. spentInputTokens: type: integer spentOutputTokens: type: integer stoppedBy: type: string nullable: true enum: - budget - aborted - cap - unpriced - failed description: >- Why it stopped short. `cap` is the per-run ceiling with no budget set, `budget` is the spend limit, `aborted` is a stop, `unpriced` means a budget was set but the provider reported no cost so the run refused to page on blind, and `failed` means every call in a page failed. Null means nothing is due any more. ReconcileRun: type: object description: One run, and what it did. properties: id: type: string format: uuid mode: type: string enum: - dry_run - apply trigger: type: string enum: - manual - idle - startup status: type: string enum: - running - success - error - aborted scanned: type: integer description: Active memories the run looked at. retired: type: integer description: Memories staled. Never deleted, and undo puts them back. folded: type: integer description: Entities folded into another. questions: type: integer description: Findings reported rather than fixed. conflictsRaised: type: integer possible: type: integer description: >- Unconfirmed contradictions the re-check recorded. Not conflicts until a human says so. llmCalls: type: integer spentUsd: type: number description: >- What the run actually cost, from the provider's own billed figures. Written per call, so it survives a crash mid-sweep. spentInputTokens: type: integer spentOutputTokens: type: integer error: type: string nullable: true description: Why a run with status `error` failed. Null otherwise. startedAt: type: string format: date-time finishedAt: type: string format: date-time nullable: true revertedAt: type: string format: date-time nullable: true ReconcileAction: type: object description: One thing a run found. properties: id: type: string format: uuid runId: type: string format: uuid kind: type: string enum: - retire - question - conflict - fold - possible description: >- `retire` and `fold` changed state. `question` and `conflict` ask you. `possible` is an unconfirmed contradiction from the re-check. class: type: string description: The detector behind it. example: duplicate_content memoryId: type: string format: uuid nullable: true reason: type: string example: identical content to 41b0e2d8, which is older and stays active applied: type: boolean description: True only when an applying run actually changed something here. surfaced: type: boolean description: False when the finding was recorded but held back by a per-run cap. decision: type: string enum: - approved - rejected - snoozed nullable: true mergeId: type: string format: uuid nullable: true description: On a fold, the merge record the undo reverses. conflictId: type: string format: uuid nullable: true description: On a conflict, the queue row this finding became. candidateId: type: string format: uuid nullable: true description: On a possible contradiction, the older belief of the pair. createdAt: type: string format: date-time ReconcileEstimate: type: object description: >- What re-checking everything currently due would cost, from the real prompt template and the actual content lengths. A dry run reports this instead of spending it. One run is capped at 200 beliefs, so on a store with a backlog this is what several runs add up to. properties: window: type: integer description: Active memories due a contradiction re-check. llmCalls: type: integer description: One per memory in the window. inputTokens: type: integer outputTokens: type: integer model: type: string example: google/gemini-2.5-flash usd: type: number nullable: true description: >- Null for a model whose rates memloom does not know, rather than a confident wrong number. EntityResolutionResult: type: object description: What one resolution pass did. properties: examined: type: integer description: Entities considered. pairs: type: integer description: Pairs that survived bucketing and were judged. merged: type: integer description: Folds applied without asking. queued: type: integer description: Uncertain folds written to the conflicts queue. deferred: type: integer description: >- Uncertain folds not written because the queue is already full. Reported rather than dropped silently, so "nothing queued" never hides "there was more to ask about". A later pass picks them up. skipped: type: integer description: Pairs already merged, queued, or settled as distinct. responses: ValidationError: description: The request body failed validation. `issues` names each offending field. content: application/json: schema: type: object properties: error: type: string example: invalid request body issues: type: array items: type: object properties: path: type: string example: query message: type: string example: query must be a non-empty string ````