openclaw

​

openclaw

​

Identity

• Native format: openclaw-agent-sqlite. One database per configured agent (~/.openclaw/agents/<agentId>/agent/openclaw-agent.sqlite, or PostgreSQL) with sessions, windows, append-only transcript events, conversations, deliveries and participants; a shared state database (state/openclaw.sqlite) with approvals, audit events, cron jobs, tasks, the delivery queue and skills. Legacy JSONL transcripts: openclaw-session-jsonl.
• Session ID: a window’s session_id (a UUID). No namespace. The logical conversation’s session_key (for example agent:main:whatsapp:group:<id>) → header conversationKey. agentId → header _meta ai.openclaw/agentId.
• Entity keys: messageId is the transcript entry id. A rewritten entry (one with supersedesEntryId) keeps the id at the root of its supersedesEntryId chain as its messageId, with supersedes naming the record it replaces and assembly {mode: rewrite, sourceIds: [<new id>]} (ASR-352, ASR-522). callId is toolCall.id; older transcripts that repeat IDs leave results unpaired with linkage: ambiguous and no callId, keeping the native ID in tool_result.nativeCallId (ASR-262). inferenceId is the assistant row’s responseId, else its turnId, else derived from the entry ID: OpenClaw’s turnId names one model response, not a turn. turnId is __openclaw.runId: OpenClaw’s run is the ASR turn.
• Runs by other harnesses: agentHarnessId names the runtime that ran the window. For the Codex app-server runtime, OpenClaw documents its transcript as a mirror of the Codex thread (“Primary thread state: Codex thread, plus OpenClaw transcript mirror”): authority: mirror and executedBy [{session: "codex:<thread id>", evidence: native_field}]. Claude Code CLI bindings (cliSessionBindings, claudeCliSessionId) and Muse bindings → executedBy without mirror, because OpenClaw records its own events for them.
• Assistant rows that are not model output: rows with provider: "openclaw" (delivery-mirror, gateway-injected, automation-result) get no inference. A delivery mirror copies a reply of the same stream: copiedFrom {record: "messageId=<sourceMessageId>", proof: native_field} and author: harness. A cron or subagent result written into the conversation → author: harness (cron) or agent (subagent), origin: injected or subagent_report.
• Who wrote a message: provenance.kind: external_user → origin: prompt, with author: human only when the sender’s identity says it is a person (senderKind: human, or a Gateway profile); inter_session → author: agent, origin: inter_session, from: {session: <originSessionId>}; internal_system → author: harness, origin: injected. custom_message entries → author: harness; their role is decided per provider at request time and is not stored: role absent with capture.role: absent. operatorMessage rows likewise have no stored role. __openclaw.modelPromptProjection → content, with the stored text in displayContent (displayCorrespondence: native). Talk rows (provenance.kind: realtime_voice) and media with transcribed: true → text parts with derivation {method: transcription}.
• Triggers: provenance.sourceTool → trigger.native. Mapping: cron or cron_execution → schedule (id from jobId, runId from runId). Use heartbeat only when evidence from the time the job ran, or a saved receipt from the job’s setup, establishes the job’s kind for that run. A current job name or uncertain conversion time does not establish the job’s kind for that run. Without that evidence, keep schedule and native provenance. Gmail and webhook hooks (hook_external_content_source) → event or webhook; main_session_restart_recovery → recovery; exec and agent_harness_completion → task_completion. A session created for a cron run → header trigger.
• Senders and conversations: __openclaw.senderId, senderName, senderUsername and senderIdentity → a participant (displayName; ids [{scheme: <channel>, value: <id>, basis: <senderIdentity.type>}]; kind from senderKind) and message.from {participant}; senderIsOwner → role: owner. conversations rows → header conversations (channel, kind, nativeId); transport.messageId and replyToId → platform; whether the agent was mentioned → addressed; transport.clients → via (Control UI → web, TUI → tui, CLI → cli, mobile → mobile, channels → channel).
• Approvals: operator_approvals → permission with requestId from the approval ID and subject.callId from source_tool_call_id. status: pending → pending, allowed and denied → resolved, expired → expired, cancelled → resolved with decision: cancelled; a terminal_reason other than user or timeout → no normalized status unless this registry defines its meaning; preserve the literal in ai.openclaw/terminalReason. decision: allow-once → allow_once, allow-always → allow_always, deny → reject_once. resolver_kind and resolver_id → decider (device → {device: resolver_id}, channel → {address: resolver_id}); no resolver ID means no actor reference. approvalReviews[] → reviews (by: {address: <label>}, verdict from status, rationale); when the automatic reviewer itself resolved the request without a person, decidedBy: classifier.
• Questions: Gateway question records are kept in memory only. The ask_user call and its result are stored → input_request with mode: questions, callId, questions and options from the input, status from the result (answered, or no_answer → timed_out) and answers.
• Tools: MCP identity is on the result: toolResult.details.mcpServer and mcpTool → tool_result.provider {kind: mcp, server, tool, identity: exact} (ASR-276); the call’s composite name stays as recorded. Code Mode nested calls (openclaw.nested-tool.v1) → tool_call with requestedBy: code and parentCallId from parentToolCallId, and its tool_result. A user-run shell command (bashExecution) → tool_call with requestedBy: user. A background command (status: running with a process ID) and the process tool → terminal with kind: shell.
• Memory: a file-tool write to the agent workspace’s MEMORY.md or memory/YYYY-MM-DD*.md (scope: agent), USER.md (scope: user) or DREAMS.md → memory_change {store: "workspace", op: add | update | delete, target.path, cause.callId}. Dreaming → memory_change with op: consolidate and cause: {actor: harness}, and a subagent with capture.childSession: absent. openclaw memory forget and memory_session_tombstones → memory_change with op: forget and target.key the forgotten session ID; status stays absent until the source records the deletion. Loaded workspace files (systemPromptReport) → instructions with kind: memory, user or project, a digest, and truncated when the file was cut before injection.
• Notes: compaction {summary, firstKeptEntryId, tokensBefore, tokensAfter} → compaction with replaces.throughId (the compaction’s native predecessor, from parentId), keptFromId and insertAt.beforeId (firstKeptEntryId), and tokens. branch_summary → kind: branch_summary. A Codex-runtime compaction without a summary → kind: provider_native with capture.summary: absent. Usage: input excludes cached tokens → input.uncached; cacheWrite1h and cacheWrite5m → cacheWriteByTtl 1h and 5m; cost.total → cost.amount in USD with source: provider_reported when totalOrigin is provider-billed, otherwise harness_estimated; a call whose usageReport.state is unavailable has no usage (not zeros); responseModel (or model when absent) → model.id, model → model.requested. model_change and thinking_level_change → config_change. Session settings (permissionMode, sandboxMode, queue mode) are current values only: write one config_change at the end of the capture, with snapshot: true and assembly: {mode: snapshot}, never backdated (ASR-034, ASR-187). SessionEntry.goal → plan with planType: goal (assembly: snapshot); the progress_card tool → plan with planType: items and markdown content. skill_workshop_changes → config_change.capability {kind: skill, name, op} (create → install, remove_file and archive → remove, other actions retained at ai.openclaw/skillAction with normalized op absent). OpenClaw records hook effects but not hook runs: no hook records. Provider compaction checkpoints are compactions, not checkpoint records; cloud workspace snapshots are checkpoint with kind: snapshot.

​

OpenClaw SQLite converter supplement

• Identity: one stream per session_windows.session_id; session_key remains conversationKey. Legacy JSONL uses header session.id. No namespace. Duplicate native entry identities are errors, not permission to merge them.
• Copies: require the recorded ancestor window and transcript_event_identities membership under the pinned window-copy contract. Matching IDs outside that contract do not prove copies. Only a contiguous prefix consisting entirely of copied records can supply inherits.localThroughId; wrapper-owned audit facts cannot be included in that prefix. Copied run IDs never establish which run or window owns facts in a side store.
• Limits: retain unknown side rows and compressed trajectory evidence without claiming their normalized semantics. Cold archives need not_retained gaps when history was lost. Unavailable context, terminal lifecycle, memory origins/tombstones, lane resolution, standing grants, boards and progress history remain declared limits where the converter cannot map them. A converter’s partial coverage does not remove the broader destinations above.

​

Position basis

• Capture order: transcript sequence, audit sequence, skill cursor_id, trajectory sequence, then mutable state. Independent database transactions are not one atomic capture. When state that spans several stores has changed, use a new capture (ASR-033/034). Stable combinations of table, primary key and projection identify records; positions based only on rowid are scoped to the capture.

​

Destinations

​

OpenClaw: native facts without a standard field

{
  "type": "ai.openclaw/reaction",
  "seq": 29,
  "id": "r29",
  "ts": "2026-10-26T08:03:51.442Z",
  "target": {"message": "messageId=b7e2d940"},
  "emoji": "👍",
  "by": {"participant": "p1"},
  "op": "add"
}

​

OpenClaw supplement: native facts without a standard field

​

Links

• Windows and lineage: OpenClaw moves a conversation to a new window for many reasons; each window is a session. session_windows.previous_session_id and reason → lineage.continues {session, reason} with the same reason names (reset, rollover, rewind, switch, recovery, compaction, and daily or idle when the window reason records them). A window with reason fork → forkedFrom, from SessionEntry.forkSource {sessionKey, sessionId, entryId} with record: "messageId=<entryId>". A new window begins with entries copied from the old one with the same IDs: inherits {mode: copied, from, localThroughId, evidence: native_structure} under the source’s copied-prefix contract. Individual copies can use proof: prefix_contract; that is not a lineage evidence value. A reset entry inside one window (reason: new | reset | idle | daily | cron-stale) → compaction with kind: reset and reason. Subagents: spawnedBy from SessionEntry.spawnedBy and subagent_runs. root may be claimed for a window with reason initial and no parent, fork or spawn field.
• Tree: native parent parentId; response key responseId, else turnId; active leaf: the last leaf.targetId, falling back to session_transcript_index_state.leaf_event_id. Injected: custom_message entries, and user rows with synthetic, runtimeContext, or a provenance.kind other than external_user. A leaf entry for a rewind → rewind {reason: user_rewind}; for a branch switch → rewind {reason: other}.

​

OpenClaw link contracts

​

Usage

​

Lane and unassigned facts

• Delivery: Assign conversation_deliveries and queue facts to a window only by an exact native window/message link or an independently owned run or queue generation. A conversationKey alone does not identify which window a fact belongs to. Unassigned facts go once in manifest _meta ai.openclaw/unassigned, with the lane’s identity (its conversation key) and provenance; declare a vendor loss only when facts are omitted, not merely unassigned. A standalone stream needs a bundle to retain these facts. When emitted, delivery.message independently requires a message reference that can be resolved (ASR-235). Map native outcomes created/queued to queued, sent/replied to sent, suppressed to suppressed, rejected to failed, and unknown to unknown; keep platform IDs, retry_count and recorded reasons. A NO_REPLY message alone is not a delivery attempt.
• Side facts and execution: copied __openclaw.runId values never establish which run or window owns facts stored outside the transcript. Without an independent original-run/window link, retain the fact as unassigned. A native null exitCode means the exit code is unavailable; omit it. Keep a numeric exitSignal as a number in raw, never in signal.

​

Losses

​

Evidence

​

Observations from 2026-10-27

{
  "name": "ai.openclaw",
  "aliases": [
    "openclaw"
  ],
  "namespace": {
    "rule": "none"
  },
  "positions": [
    {
      "family": "transcript",
      "basis": "native-sequence",
      "boundary": "transaction"
    },
    {
      "family": "audit",
      "basis": "native-sequence",
      "boundary": "transaction"
    },
    {
      "family": "current-state",
      "basis": "snapshot",
      "boundary": "transaction"
    }
  ],
  "formats": [
    {
      "id": "openclaw-agent-sqlite",
      "evidence": [
        0,
        3
      ]
    }
  ],
  "evidence": [
    {
      "label": "source",
      "revision": "2038c84",
      "scope": "format-and-explicit-mappings"
    },
    {
      "label": "authored",
      "scope": "converter-tests-only"
    },
    {
      "label": "probe",
      "version": "2026.10.7",
      "scope": "persistence-only"
    },
    {
      "label": "observed",
      "version": "2026.10.8 (dc2523e)",
      "date": "2026-10-27",
      "scope": "agent SQLite schema 25 and shared SQLite schema 20; Edit, shell test, resume, review/question; 2 streams and a native input_request.",
      "limitations": "OpenClaw 2026.10.8 (dc2523e) wrote agent schema 25 and shared schema 20. Qualify storage schemas separately from release numbers. For shared skill/cron tables without cursor_id, the observed order is timestamp then rowid. Edit, shell-test, resume and review/question runs retained a native input_request. Preserve that request without inventing an answer or approval.",
      "validation": {
        "streams": 2,
        "records": 132,
        "errors": 0,
        "notices": 0,
        "gaps": 0,
        "unidentifiedSources": 0
      },
      "validator": "validator-v0.1.2",
      "converter": "ctx-asr-authentic-captures-4"
    }
  ],
  "destinations": {
    "agentId": {
      "schema": {},
      "evidence": [
        0
      ]
    },
    "appendMode": {
      "schema": {},
      "evidence": [
        0
      ]
    },
    "async": {
      "schema": {},
      "evidence": [
        0
      ]
    },
    "compaction": {
      "schema": {},
      "evidence": [
        0
      ]
    },
    "consumedAt": {
      "schema": {},
      "evidence": [
        0
      ]
    },
    "consumedBy": {
      "schema": {},
      "evidence": [
        0
      ]
    },
    "consumerId": {
      "schema": {},
      "evidence": [
        0
      ]
    },
    "contextUsage": {
      "schema": {},
      "evidence": [
        0
      ]
    },
    "costComponents": {
      "schema": {
        "type": "object"
      },
      "evidence": [
        0
      ]
    },
    "decisionFacts": {
      "schema": {},
      "evidence": [
        0
      ]
    },
    "delivery": {
      "schema": {},
      "evidence": [
        0
      ]
    },
    "executionMode": {
      "schema": {},
      "evidence": [
        0
      ]
    },
    "expiresAt": {
      "schema": {},
      "evidence": [
        0
      ]
    },
    "goal": {
      "schema": {
        "type": "object"
      },
      "evidence": [
        0
      ]
    },
    "goalState": {
      "schema": {
        "type": "object"
      },
      "evidence": [
        0
      ]
    },
    "harnessId": {
      "schema": {},
      "evidence": [
        0
      ]
    },
    "humanMentions": {
      "schema": {},
      "evidence": [
        0
      ]
    },
    "identity": {
      "schema": {},
      "evidence": [
        0
      ]
    },
    "media": {
      "schema": {},
      "evidence": [
        0
      ]
    },
    "message": {
      "schema": {
        "type": "object"
      },
      "evidence": [
        0
      ]
    },
    "mirror": {
      "schema": {},
      "evidence": [
        0
      ]
    },
    "node": {
      "schema": {
        "type": "object"
      },
      "evidence": [
        0
      ]
    },
    "operationId": {
      "schema": {},
      "evidence": [
        0
      ]
    },
    "parentToolCallId": {
      "schema": {},
      "evidence": [
        0
      ]
    },
    "presented": {
      "schema": {},
      "evidence": [
        0
      ]
    },
    "providerReplay": {
      "schema": {},
      "evidence": [
        0
      ]
    },
    "queueId": {
      "schema": {},
      "evidence": [
        0
      ]
    },
    "replyToPreview": {
      "schema": {},
      "evidence": [
        0
      ]
    },
    "requestedAt": {
      "schema": {},
      "evidence": [
        0
      ]
    },
    "requesterOrigin": {
      "schema": {},
      "evidence": [
        0
      ]
    },
    "resolverKind": {
      "schema": {},
      "evidence": [
        0
      ]
    },
    "risk": {
      "schema": {},
      "evidence": [
        0
      ]
    },
    "role": {
      "schema": {},
      "evidence": [
        0
      ]
    },
    "senderIsOwner": {
      "schema": {},
      "evidence": [
        0
      ]
    },
    "skillAction": {
      "schema": {},
      "evidence": [
        0
      ]
    },
    "skillChange": {
      "schema": {},
      "evidence": [
        0
      ]
    },
    "sourceSessions": {
      "schema": {},
      "evidence": [
        0
      ]
    },
    "standingGrant": {
      "schema": {},
      "evidence": [
        0
      ]
    },
    "subagent": {
      "schema": {},
      "evidence": [
        0
      ]
    },
    "submittedBy": {
      "schema": {
        "type": "object"
      },
      "evidence": [
        0
      ]
    },
    "terminalReason": {
      "schema": {},
      "evidence": [
        0
      ]
    },
    "thoughtSignature": {
      "schema": {},
      "evidence": [
        0
      ]
    },
    "transcriptVersion": {
      "schema": {},
      "evidence": [
        0
      ]
    },
    "usageReport": {
      "schema": {},
      "evidence": [
        0
      ]
    },
    "window": {
      "schema": {
        "type": "object"
      },
      "evidence": [
        0
      ]
    }
  },
  "mappings": [
    {
      "native": "complete native row or event",
      "destination": "native.record or native.blob",
      "scope": "agent SQLite schema 25 and shared SQLite schema 20",
      "evidence": [
        3
      ]
    }
  ]
}