Session management deep dive

A single Gateway process owns session state end-to-end. UIs (macOS app, web Control UI, TUI) query the Gateway for session lists and token counts. In remote mode, the per-agent SQLite database lives on the remote host, so checking your local Mac’s state will not reflect what the Gateway is using.

Overview docs first: Session management, Compaction, Memory overview, Memory search, Session pruning, Transcript hygiene, full config reference at Agent config.

Two persistence layers

1. Session rows (per-agent SQLite) - key/value map sessionKey -> SessionEntry. Mutable runtime state owned by the Gateway. Tracks metadata: current session id, last activity, toggles, token counters.
2. Transcript events (per-agent SQLite) - append-only, tree-structured (entries have id + parentId). Stores the conversation, tool calls, and compaction summaries; rebuilds model context for future turns. Compaction checkpoints are metadata over the compacted successor transcript - a new compaction does not write a second .checkpoint.*.jsonl copy.

Older installs may still have sessions.json files under the agent sessions/ directory. Treat those files as legacy session-row migration inputs or explicit offline-maintenance targets. Gateway startup does not import them. Stop the Gateway, back up its state, and use openclaw doctor --fix to import legacy rows and transcript history into the per-agent SQLite store. Run openclaw doctor --session-sqlite inspect --session-sqlite-all-agents, then follow the Doctor migration sequence for inspection and validation. If a migration fails after legacy transcript artifacts were archived, use the Doctor recovery mode from that sequence. Recovery uses migration manifests, restores only the affected archived support artifacts, prepares a sanitized GitHub issue report when requested, and does not make active runtime read JSONL files again.

Gateway history readers avoid materializing the whole transcript unless the surface needs arbitrary historical access. First-page history, embedded chat history, restart recovery, and token/usage checks use bounded tail reads from SQLite. Full transcript scans go through the async transcript index and are shared across concurrent readers.

On-disk locations

Per agent, on the Gateway host (resolved via src/config/sessions.ts):

• Runtime session row store: ~/.openclaw/agents/<agentId>/agent/openclaw-agent.sqlite
• Runtime transcript rows: ~/.openclaw/agents/<agentId>/agent/openclaw-agent.sqlite
• Legacy/archive transcript artifacts: ~/.openclaw/agents/<agentId>/sessions/
• Legacy row migration input: ~/.openclaw/agents/<agentId>/sessions/sessions.json

Store maintenance and disk controls

session.maintenance controls automatic maintenance for SQLite session rows, SQLite transcript rows, archive artifacts, and trajectory sidecars:

Key	Default	Notes
mode	"enforce"	or "warn" (report only, no mutation)
pruneAfter	"30d"	stale-entry age cutoff
archiveDashboardAfter	"7d"	inactivity cutoff for dashboard sessions; false or 0 disables
maxEntries	500	cap on unarchived session rows when protection permits
resetArchiveRetention	keep (no age cutoff)	age cutoff for *.reset.*/*.deleted.* transcript archives; a duration opts into deletion
maxDiskBytes	10gb	per-agent sessions disk budget; false, 0, or "0" disables
highWaterBytes	80% of maxDiskBytes	target after cleanup; zero-resolving values use the default, and negatives are invalid

Reset boundaries start a fresh history window without deleting earlier transcript rows. When session rollover advances the live sessionKey -> sessionId mapping, the previous SQLite session, transcript, trajectory, and search rows also remain; ordinary entry and session lists show only the live mapping. Retained reset history is bounded by the disk budget, not by resetArchiveRetention, which only ages archive artifacts. Explicit deletion is different: it stores and verifies the compressed transcript archive in SQLite in the same transaction that removes the deleted session’s rows. It then publishes, syncs, and reads back the derived *.jsonl.deleted.<timestamp>.zst file before reporting success when zstd is available.

maxDiskBytes enforcement uses physical bytes: the per-agent SQLite main file, its -wal file, and counted files in the agent sessions directory. It never estimates row JSON sizes or subtracts logical row sizes from that total.

Gateway model-run probe sessions (keys matching agent:*:explicit:model-run-<uuid>) get a separate, fixed 24h retention. This pruning is pressure-gated: it only runs when session-entry maintenance/cap pressure is reached, and only before the global stale-entry cleanup/cap step. Other explicit sessions do not use this retention.

When combined physical usage exceeds maxDiskBytes, mode: "enforce" first reclaims checkpointable database space, then removes the oldest retained reset/delete archives. If usage is still above highWaterBytes, it walks historical SQLite sessions by sessions.updated_at, oldest first. Historical means the session id is not referenced by a live session entry, a route target, or an admitted/in-flight run. For each victim, cleanup stores the compressed archive in the same write transaction that removes the session row and its transcript, trajectory, active, index, and FTS projections. It publishes, syncs, and reads back the derived file after commit. This includes sessions that contain trajectory events but no transcript events. If those tiers are insufficient, cleanup permanently deletes the oldest sessions whose recorded archive reason is active-session-cap. Manual, legacy, stale-dashboard, and recovery archives remain protected. Cleanup rechecks entry identity and admission references at deletion time, remeasures physical usage after each victim, and stops at highWaterBytes.

Committed writes and deletion first land in the WAL. Cleanup checkpoints it so the WAL can shrink immediately, then uses incremental vacuum to return eligible free tail pages from the main file; pages that are not yet reclaimable stay in the main file and therefore remain counted on the next physical measurement. mode: "warn" reports the current physical overage without checkpointing, writing an archive, or deleting rows.

Run maintenance on demand:

openclaw sessions cleanup --dry-run
openclaw sessions cleanup --enforce

maxEntries counts unarchived session rows; archived rows do not consume the cap. Cleanup archives the oldest eligible ordinary sessions until the unarchived total reaches maxEntries or no eligible victims remain. Pinned sessions, active or admitted work, model-locked sessions, and durable external conversation pointers such as group sessions and thread-scoped chat sessions remain protected, so protected rows can keep the unarchived total above the cap. Synthetic runtime entries (cron, hooks, heartbeat, ACP, sub-agents) remain disposable and can still be removed once they exceed the configured age, count, or disk budget. Isolated cron runs use a separate cron.sessionRetention control, independent of model-run probe retention.

Every new archive records a structured reason automatically. Explicit archive actions record manual; count-cap and stale-dashboard maintenance record their respective causes; recovery archives record restart-recovery. The Control UI renders a human-readable explanation. Missing or unrecognized reasons are treated as protected legacy state rather than inferred.

--dry-run previews the unarchived-row cap and identifies the unprotected rows that would satisfy it; --enforce applies that cleanup immediately but does not remove protection. To reduce protected history, unarchive, unpin, wait for active work to finish, or explicitly delete sessions you no longer want to retain.

Normal Gateway writes flow through the session accessor, which serializes per-agent SQLite mutations through the runtime writer path. Runtime code should prefer the accessor helpers in src/config/sessions/session-accessor.ts; legacy sessions.json helpers are migration and offline-maintenance tools. When a Gateway is reachable, non-dry-run openclaw sessions cleanup and openclaw agents delete delegate store mutations to the Gateway so cleanup joins the same writer queue; --store <path> is the explicit offline repair path for a selected legacy store and always stays local (as does --dry-run). maxEntries cleanup is batched for production-sized stores, so the unarchived population may briefly exceed the configured cap before the next high-water cleanup rewrites it down. Reads never prune or cap entries during Gateway startup - only writes or openclaw sessions cleanup --enforce do, and the latter also applies the cap immediately and prunes old unreferenced legacy transcript, checkpoint, and trajectory artifacts even with no disk budget configured.

OpenClaw no longer creates automatic sessions.json.bak.* rotation backups during Gateway writes. The current schema rejects the legacy session.maintenance.rotateBytes key, and openclaw doctor --fix removes it from older configs.

Migration recovery originals and exact pre-Doctor recovery files are separate from ordinary session retention: they are excluded from the live session disk budget and have no automatic expiration. After verifying the upgrade, use openclaw update cleanup --dry-run to inspect them online. Explicit offline update cleanup can retire verified originals without removing current SQLite history; exclusion from the disk budget is not deletion authority.

Transcript mutations pass through the session accessor and SQLite writer queue. Each mutation verifies the active run’s durable writer claim inside its commit transaction, so a superseded run cannot write to the transcript.

Downgrading After The SQLite Flip

Stop the Gateway and back up its state. Using the current SQLite-capable OpenClaw version, restore archived legacy session stores and transcript artifacts before starting an older file-backed version:

openclaw doctor --session-sqlite restore --session-sqlite-all-agents

The migration archives imported hot transcript JSONL files and verified, fully covered legacy sessions.json stores in session-sqlite-import-archive/. Legacy stores with incomplete coverage or blocking migration issues remain in place. Older file-backed runtimes need both sessions.json and the artifacts referenced by its sessionFile paths at their original locations before startup.

Restore uses migration manifests, moves only recorded archived artifacts whose original paths are missing, reports conflicts rather than overwriting existing files, and leaves the SQLite database in place for forward recovery.

Originals retired by openclaw update cleanup can no longer be restored from the migration archive. Restore reports intentional disposal or pending cleanup instead of treating either as an unexpectedly missing file. An independent backup containing the legacy artifacts is required if you need them after disposal; see Pre-update backups.

Restore does not export changes made only in SQLite after migration. Sessions created after the SQLite flip are SQLite-only and will not appear to an older file-backed runtime. If you re-upgrade after a downgrade, run the Doctor inspection and validation sequence again so OpenClaw can verify restored legacy artifacts before importing.

Cron sessions and run logs

Isolated cron runs create their own session entries/transcripts with dedicated retention:

• cron.sessionRetention (default "24h") prunes old isolated cron run sessions from the store; false or a zero duration such as "0h" disables.
• Terminal run history is retained for 7 days (lost rows for 24 hours), with the newest 2000 rows per job and history class enforced as an additional ceiling.

When cron force-creates a new isolated run session, it sanitizes the previous cron:<jobId> session entry before writing the new row: it carries safe preferences (thinking/fast/verbose/reasoning settings, labels, display name) and explicit user-selected model/auth overrides, but drops ambient conversation context (channel/group routing, send/queue policy, elevation, origin, ACP runtime binding) so a fresh isolated run cannot inherit stale delivery or runtime authority from an older run.

Session keys (sessionKey)

A sessionKey identifies which conversation bucket you are in (routing + isolation). Canonical rules: /concepts/session.

Pattern	Example
Main/direct chat (per agent)	agent:<agentId>:main
Group	agent:<agentId>:<channel>:group:<id>
Room/channel (Discord/Slack)	agent:<agentId>:<channel>:channel:<id> or ...:room:<id>
Cron	cron:<job.id>
Webhook	hook:<uuid> (unless overridden)

Session ids (sessionId)

Each sessionKey points at a current sessionId (the SQLite transcript identity that continues the conversation). Decision logic lives in initSessionState() in src/auto-reply/reply/session.ts.

• Gateway reset (/new, /reset) records a reset boundary in an existing persisted session and keeps its sessionId. A session that does not exist yet receives a new id.
• No automatic reset is the default. The current sessionId continues while compaction keeps the active model context bounded.
• Daily reset (session.reset.mode: "daily") creates a new sessionId on the next message after the configured local-hour boundary (session.reset.atHour, default 4).
• Idle expiry (session.reset.mode: "idle" with session.reset.idleMinutes, or legacy session.idleMinutes) creates a new sessionId when a message arrives after the idle window. If daily and idle are both configured, whichever expires first wins.
• Control UI reconnect resume preserves the currently visible session for one reconnect send when the Gateway receives the matching sessionId from an operator UI client. This is a one-shot signal; ordinary stale sends still create a new sessionId.
• System events (heartbeat, cron wakeups, exec notifications, gateway bookkeeping) may mutate the session row but never extend daily/idle reset freshness. Reset rollover discards queued system-event notices for the previous session before the fresh prompt is built.
• Automatic parent fork policy uses OpenClaw’s active branch when creating a thread or subagent fork. If that branch is too large (over a fixed internal cap, currently 100K tokens), OpenClaw starts the child with isolated context instead of failing or inheriting unusable history. Sizing is automatic and not configurable; legacy session.parentForkMaxTokens config is removed by openclaw doctor --fix.
• Operator forks: sessions.create { parentSessionKey, fork: true } branches from the parent’s current state. Admission uses the selected child model’s effective usable input capacity, falling back to the 100K safety cap when model capacity is unavailable. A normal fork is refused while the parent has an active run; adding forkFrom: "last-completed" copies only through the last completed assistant message, excluding the in-progress tail. Unlike automatic parent forks, an operator fork over its capacity limit is rejected rather than accepted with isolated context. The child inherits the parent’s model selection unless one is passed explicitly. The response marks it forkedFromParent, and token counters start fresh.
• Message forks: sessions.fork { sessionKey, entryId } creates a child from the active-path prefix before the selected user message and returns that message to the composer for editing. The parent remains unchanged. Incognito forks retain the parent’s in-memory storage class; restarting the Gateway removes both sessions. Codex fork verification compares complete attested submitted prompts, including whitespace; the bounded display-import projection is not a substitute for that evidence. See Control UI for fork and rewind actions.

Session store schema

The runtime store keeps SessionEntry values in per-agent SQLite. The value type is SessionEntry in src/config/sessions.ts. Key fields (not exhaustive):

• sessionId: current transcript id used to address SQLite transcript rows
• sessionStartedAt: start timestamp for the current sessionId; daily reset freshness uses this. Legacy rows may derive it from the JSONL session header.
• lastInteractionAt: last real user/channel interaction timestamp; idle reset freshness uses this so heartbeat, cron, and exec events do not keep sessions alive. Legacy rows without this field fall back to the recovered session start time.
• updatedAt: last store-row mutation timestamp, used for listing/pruning/bookkeeping - not the daily/idle freshness authority.
• archivedAt: optional archive timestamp. Archived sessions stay in the store with their transcript intact and are excluded from normal active listings.
• pinnedAt: optional pin timestamp. Active pinned sessions sort ahead of unpinned sessions; archiving a session clears its pin.
• Codex thread interop: both fields follow the Codex thread-management shape - the archived/pinned booleans on the wire are always derived from the timestamp and stamped server-side, matching Codex threads.archived_at semantics and camelCase serialization. OpenClaw timestamps are epoch milliseconds while Codex uses epoch seconds, so bridges convert at the codex plugin seam. Codex has no pin API yet (thread/archive/thread/unarchive only); pinned state stays OpenClaw-side until one exists, at which point the matching shape lets bound sessions round-trip pin state mechanically.
• Codex supervision lists only non-archived native threads. A Gateway-local idle or notLoaded activity-unknown thread can be archived through native thread/archive only after the operator explicitly confirms that no other Codex process owns it; the plugin performs a fresh process-local status read first, and the thread then disappears from the catalog. That read cannot prove that another App Server process is not using the thread. OpenClaw refuses to archive active and error rows, and paired-node archive is unavailable until the node bridge can own the full streamed thread lifecycle. Unarchiving in a native Codex client makes the thread eligible to appear again.
• lastReadAt / markedUnreadAt: read-state timestamps stamped server-side by sessions.patch { unread } - unread: false records a read (sets lastReadAt, clears markedUnreadAt); unread: true records markedUnreadAt and marks the session unread until the next activation or explicit read. Session rows expose the marker alongside a derived unread boolean so already-open clients preserve manual reminders while still acknowledging new activity. Automatic read patches from clients that support the advertised unread acknowledgement contract include expectedMarkedUnreadAt (null means no marker); a newer marker makes that acknowledgement a successful no-op instead of erasing newer intent. Bare unread: false requests retain the legacy clear behavior, so protection across several connected clients requires each active client to support the contract. Sessions never marked read stay unread: false, so existing installs do not light up on upgrade.
• lastActivityAt: timestamp of the last completed agent run that counts as unread-worthy activity (user, channel, and cron runs). Heartbeat and internal-event turns, plus metadata patches, do not update it; updatedAt is not an activity signal.
• sessionFile: legacy marker retained for migration/archive compatibility; active runtime uses SQLite identity
• chatType: direct | group | room
• provider, subject, room, space, displayName: group/channel labeling metadata
• Toggles: thinkingLevel, verboseLevel, reasoningLevel, elevatedLevel, sendPolicy (per-session override)
• Model selection: providerOverride, modelOverride, authProfileOverride
• Token counters (best-effort/provider-dependent): inputTokens, outputTokens, totalTokens, contextTokens
• compactionCount: how many times auto-compaction completed for this session key
• memoryFlushAt / memoryFlushCompactionCount: timestamp and compaction count of the last pre-compaction memory flush

The Gateway is the authority: it may rewrite or rehydrate entries as sessions run. For legacy file-backed installs, migrate with openclaw doctor --session-sqlite import --session-sqlite-all-agents instead of editing sessions.json and expecting runtime to keep reading that file.

Transcript event structure

Transcripts are managed by the OpenClaw session accessor and exposed to runtime code through identity-based helpers. The event stream is append-only:

• First entry: session header - type: "session", id, cwd, timestamp, optional parentSession.
• Then: entries with id + parentId (tree structure).

Notable entry types:

• message: user/assistant/toolResult messages
• custom_message: extension-injected message that does enter model context (rendered in the TUI when display: true, hidden entirely when display: false)
• custom: extension state that does not enter model context (for persisting extension state across reloads)
• compaction: persisted compaction summary with firstKeptEntryId and tokensBefore
• reset: a fresh history window, optionally retaining messages from firstKeptEntryId
• branch_summary: persisted summary when navigating a tree branch

History readers keep the latest reset window across later compactions: explicitly retained reset messages and messages after that reset remain visible, but older messages and compaction summaries do not reappear. Model context follows the latest reset or compaction instead, so compaction can summarize the current conversation without reopening its earlier history.

Model-only callers should await SessionManager.openModelContextAsync(target, { admission?, signal? }) to create a detached, non-persisting view without blocking the Gateway event loop on durable history scans. openModelContext() provides the same view for synchronous consumers. The reader selects payloads in SQLite and retains lightweight navigation outside the model window, without introducing a history size cutoff. Storage-only native prompt text and tool-result details stay out of that view; mirror identity, sender and media facts, tool content, and valid provider replay state remain available. Native fork verification, replay, exports, and doctor operations continue to use full-fidelity evidence readers.

SessionManager.readSessionContext(target, read, { admission? }) lets a synchronous consumer process full-fidelity context messages inside one read-only snapshot. The callback receives (messages, header): a lazy message iterable and the unvalidated stored header. Missing stores supply an empty iterable and no header without creating a database. An optional admission excludes the current admitted user row and later events. Callback errors propagate, and the iterator closes when the callback returns or throws; it cannot be retained for later reads. This lets replay consumers enforce their existing limits during acquisition without silently dropping earlier history. Navigation still scales with the transcript, and individual selected rows are decoded whole; this is not a fixed process-memory ceiling.

Durable model-context reads run in a worker. Codex native replay and settled-turn verification keep that lazy read and its consumer together in a worker. Worker reads are serialized per reader and reuse an idle worker. Admission receipts are validated inside the read snapshot and again before the result is accepted; reads without an admission instead check the session’s rewrite generation and last event sequence. Admitted reads retain their turn boundary, so later appends alone do not invalidate them. An invalidated read or canceled signal rejects the result. Callers carry their original cancellation signal through context acquisition and check that their owner remains active before invoking hooks, starting a model run, or applying a proposal. Incognito sessions use the same operation in the Gateway process because their SQLite database is held in memory.

OpenClaw intentionally does not “fix up” transcripts; the Gateway uses SessionManager to read/write them.

Context windows vs tracked tokens

Two different concepts:

1. Model context window: hard cap per model (tokens visible to the model). Comes from the model catalog and can be overridden via config.
2. Session store counters: rolling stats written into the session row (used for /status and dashboards). contextTokens is a runtime estimate/reporting value - do not treat it as a strict guarantee.

Completed turns update session counters even when no compaction occurred. An ordered context replacement takes precedence over earlier model usage; if its size is unknown, the counters are marked stale instead of borrowing an older request’s total. A superseded run cannot overwrite the current writer’s counters.

More on limits: /reference/token-use.

Compaction: what it is

Compaction summarizes older conversation into a persisted compaction entry in the transcript and keeps recent messages intact. After compaction, future turns see the compaction summary plus messages after firstKeptEntryId. Compaction is persistent, unlike session pruning - see /concepts/session-pruning.

Embedded OpenClaw compaction uses low thinking by default. Set agents.defaults.compaction.thinkingLevel: "inherit" to reuse the session level, or choose another explicit level for summary calls; the runtime clamps it to each concrete compaction model or fallback. Native Codex app-server compaction owns its compact request and cannot accept a per-compaction thinking override, so OpenClaw warns and leaves that setting to Codex.

AGENTS.md section reinjection after compaction remains opt-in via agents.defaults.compaction.postCompactionSections. Plugins can add other prompt context through before_prompt_build.

Chunk boundaries and tool pairing

When splitting a long transcript into compaction chunks, OpenClaw keeps assistant tool calls paired with their matching toolResult entries:

• If the token-share split would land between a tool call and its result, OpenClaw shifts the boundary to the assistant tool-call message instead of separating the pair.
• If a trailing tool-result block would otherwise push the chunk over target, OpenClaw preserves that pending tool block and keeps the unsummarized tail intact.
• Aborted/error tool-call blocks do not hold a pending split open.

When auto-compaction happens

The built-in OpenClaw runtime has three scheduling paths:

1. Overflow recovery: the model returns a context-overflow error (request_too_large, context length exceeded, input exceeds the maximum number of tokens, input token count exceeds the maximum number of input tokens, input is too long for the model, ollama error: context length exceeded, and other provider-shaped variants) - compact, then retry. When the provider reports the attempted token count, OpenClaw forwards that observed count into overflow-recovery compaction; if the provider confirms overflow but exposes no parseable count, OpenClaw passes a minimally over-budget synthetic count to compaction engines and diagnostics. If overflow recovery still fails, OpenClaw surfaces explicit guidance and preserves the current session mapping instead of silently rotating to a fresh session id - retry the message, run /compact, or run /new.

   One provider shape is terminal rather than compaction-recoverable. When the refusal states a single request larger than the provider’s entire token limit - Groq answers an oversized request with an HTTP 413 naming TPM and stating Limit <n>, Requested <m> - no bucket state can admit it. Compaction budgets against the model’s context window rather than that per-request ceiling, and its own summarization request is refused by the same ceiling, so it can only spend further calls that cannot succeed. OpenClaw surfaces the reset guidance immediately instead of compacting, adopting a successor transcript, or retrying. Ordinary TPM throttling, which states a requested size within the limit, stays a rate limit and keeps its normal backoff.

2. Usage-based maintenance: normal replies check projected usage before their turn; successful direct commands, including agent --local and Gateway agent RPC runs, check after the completed turn is persisted and any pending final reply is protected. Both block on projected usage at or above the active model window minus the selected compaction reserve, subject to an applicable server compaction threshold floor. The memory-flush soft margin does not lower this blocking threshold. Disabling memory flush does not disable this compaction. Direct-command maintenance respects compaction.enabled: false and skips a second compaction when the completed run already compacted.

3. Session-internal threshold maintenance: default-mode sessions can also compact when actual context usage exceeds the model window minus the session reserve. Safeguard mode disables this competing session-internal path and leaves proactive scheduling to the maintenance owner above.

The persisted contextBudgetStatus is a pre-prompt pressure estimate, not an execution command. Completed direct commands, normal replies, and queued follow-up replies record it when the runtime supplies one. /status can show this estimate, marked with ~ and est, when fresh token usage is unavailable. Compaction and session resets invalidate old estimates; a completed run without a diagnostic clears the previous one unless that run preserves the session’s model state (for example, a heartbeat). Its route and shouldCompact fields can report pressure while the provider attempt is still admitted. Use completed compaction counts and transcript entries to verify that compaction actually happened.

Two additional guards run outside these paths:

• Preflight local compaction: set agents.defaults.compaction.maxActiveTranscriptBytes to a positive byte threshold (bytes or a string like "20mb") to trigger local compaction before opening the next run once the active transcript reaches that size. Normal semantic compaction still runs. For Codex app-server sessions, the same threshold caps native rollout transcripts and oversized native threads restart fresh. Unset or 0 disables the guard.
• Mid-turn precheck: set agents.defaults.compaction.midTurnPrecheck.enabled: true (default false) to add a tool-loop guard. After a tool result is appended and before the next model call, OpenClaw estimates prompt pressure using the same preflight budget logic used at turn start. If context no longer fits, the guard does not compact inline - it raises a structured mid-turn precheck signal, stops the current prompt submission, and lets the outer run loop use the existing recovery path (truncate oversized tool results when that is enough, or trigger the configured compaction mode and retry). Works with both default and safeguard compaction modes, including provider-backed safeguard compaction. Independent of maxActiveTranscriptBytes: the byte-size guard runs before a turn opens, mid-turn precheck runs later, after new tool results are appended.

Compaction settings

{
  agents: {
    defaults: {
      compaction: {
        enabled: true,
        keepRecentTokens: 20000,
      },
    },
  },
}

OpenClaw enforces a built-in reserve for embedded runs and caps it against the active model context window so it cannot consume the whole prompt budget. This keeps small-context local models from entering compaction from the first token while leaving enough headroom for multi-turn housekeeping such as the memory flush.

Set enabled: false to disable threshold-driven auto-compaction inside the embedded agent runtime and direct-command post-turn maintenance. OpenClaw’s reply preflight and overflow-recovery compaction paths remain available, and manual /compact continues to work.

Manual /compact uses agents.defaults.compaction.keepRecentTokens (default: 20000) and keeps that recent-tail cut point.

OpenClaw adopts an explicit successor identity returned by a context engine. The built-in SQLite compactor keeps the current session identity. Branch/restore checkpoint actions use a returned successor when present; legacy pre-compaction checkpoint files remain readable while referenced.

Pluggable compaction providers

Plugins register a compaction provider via registerCompactionProvider() on the plugin API. When agents.defaults.compaction.provider is set to a registered provider id, the safeguard extension delegates summarization to that provider instead of the built-in summarizeInStages pipeline.

• provider: id of a registered compaction provider plugin. Leave unset for default LLM summarization. Setting a provider forces mode: "safeguard".
• Providers receive the same compaction instructions and identifier-preservation policy as the built-in path, and the safeguard still preserves recent-turn and split-turn suffix context after provider output.
• Built-in safeguard summarization re-distills prior summaries with new messages instead of preserving the full previous summary verbatim.
• Safeguard mode enables built-in summary quality audits by default. After final budgeting, the retained generated body must contain the required headings, and the exact artifact to be persisted must retain pending asks and exact identifiers. Corrective attempts stay within qualityGuard.maxRetries; exhaustion or a corrective generation failure cancels before append and leaves the original transcript authoritative. Set qualityGuard.enabled: false to skip this behavior. Configured compaction-provider output remains outside the built-in audit loop.
• If the provider fails or returns an empty result, OpenClaw falls back to built-in LLM summarization automatically. Provider-local failures, including timeouts, stay in that guarded fallback and use the built-in quality audit when enabled. Abort/timeout signals the caller explicitly triggered are re-thrown, not swallowed, so cancellation is always respected.

Source: src/plugins/compaction-provider.ts, src/agents/agent-hooks/compaction-safeguard.ts.

User-visible surfaces

• /status in any chat session
• openclaw status (CLI)
• openclaw sessions / openclaw sessions --json
• Gateway logs (pnpm gateway:watch or openclaw logs --follow): embedded run auto-compaction start + complete
• Verbose mode: 🧹 Auto-compaction complete plus the compaction count

Silent housekeeping (NO_REPLY)

OpenClaw supports “silent” turns for background tasks where the user should not see intermediate output.

• The assistant starts its output with the exact silent token NO_REPLY / no_reply to mean “do not deliver a reply to the user.” OpenClaw strips/suppresses this in the delivery layer.
• Exact silent-token suppression is case-insensitive: NO_REPLY and no_reply both count when the whole payload is just the silent token.
• As of 2026.1.10, OpenClaw also suppresses draft/typing streaming when a partial chunk begins with NO_REPLY, so silent operations do not leak partial output mid-turn.
• This is for true background/no-delivery turns only - it is not a shortcut for ordinary actionable user requests.

Pre-compaction memory flush

Before auto-compaction happens, OpenClaw can run a silent agentic turn that writes durable state to disk (for example memory/YYYY-MM-DD.md in the agent workspace) so compaction cannot erase critical context. It monitors session context usage, and once it crosses a soft threshold below the compaction threshold, it sends a silent “write memory now” directive using the exact silent token NO_REPLY / no_reply so the user sees nothing.

Config (agents.defaults.compaction.memoryFlush), full reference at /gateway/config-agents:

Key	Default	Notes
enabled	true
model	unset	exact provider/model override for the flush turn only, for example ollama/qwen3:8b
softThresholdTokens	4000	gap below the compaction threshold that triggers a flush
forceFlushTranscriptBytes	unset (disabled)	force a flush once active transcript history reaches this estimated byte size (or string like "2mb"), even if token counters are stale; 0 disables

For a 32,768-token window, a 20,000-token reserve and a 4,000-token soft margin, early flushing starts at 8,768 projected tokens. Blocking token compaction starts at 12,768, or later if an applicable server threshold is higher. Between those thresholds, flushing can run without blocking the next user turn on compaction. The selected memory provider owns the reserve and flush margin; without a flush plan, maintenance still uses the effective compaction reserve. Nonpositive thresholds suppress token triggers. Transcript byte guards remain independent.

Notes:

• The built-in prompt and system prompt include a NO_REPLY hint to suppress delivery.
• When model is set, the flush turn uses that model without inheriting the active session’s fallback chain, so local-only housekeeping does not silently fall back to a paid conversation model on failure.
• The flush runs once per compaction cycle (tracked in the session row).
• The flush runs only for embedded OpenClaw sessions; CLI backends and heartbeat turns skip it.
• The flush is skipped when the session workspace is read-only (workspaceAccess: "ro" or "none").
• See Memory for the workspace file layout and write patterns.

OpenClaw exposes a session_before_compact hook in the extension API, but the flush logic above lives on the Gateway side (src/auto-reply/reply/memory-flush.ts, src/auto-reply/reply/agent-runner-memory.ts), not on that hook.

Troubleshooting checklist

• Session key wrong? Start with /concepts/session and confirm the sessionKey in /status.
• Store vs transcript mismatch? Confirm the Gateway host and the store path from openclaw status.
• Compaction spam? Check the model’s context window (too small forces frequent compaction) and tool-result bloat (tune session pruning).
• Every prompt seems to overflow on a small local model? Confirm the provider reports the correct model context window. OpenClaw can cap the effective reserve only when that window is known.
• Silent turns leaking? Confirm the reply starts with the exact silent token NO_REPLY (case-insensitive) and you are on a build that includes the streaming-suppression fix (2026.1.10+).

Related

• Session management
• Session pruning
• Context engine
• Agent config reference