# Session state on disk

> Where the Gateway keeps session rows and transcripts, and the on-disk paths per agent

- 网址：https://funcoding.ai/agents/openclaw/reference/session-management-compaction/store/
- 来源：OpenClaw 官方文档原文（英文），MIT 许可，同步于 2026-10-11
- 官方原文：https://docs.openclaw.ai/zh-CN/reference/session-management-compaction/store

---
## 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 summaries and available token measurements remain in the transcript without a separate checkpoint record or snapshot copy.

Guarded transcript turns retain the selected stored key and session window. A
qualified identity such as `agent:main:global` can append to its existing raw
`global` row without renaming it or creating another row. If both spellings
already exist, the write refuses the ambiguous selection; exact reads keep the
two stored addresses separate. This also applies to retained history windows.

New chat turns use a qualified admission identity while retaining the selected
stored key and database. If a raw `global` or `unknown` row and its qualified
spelling would share that identity, chat execution refuses the ambiguous
selection, including when the raw row has not been created yet. Choose an
unambiguous session; the existing rows and history are not merged or renamed.
Completed-send retries still return their recorded result. Read-only history and
stored parent, spawn, and fork references keep their physical addresses.
Input persistence and queued-input cleanup stay bound to the admitted database;
replacing its path cannot transfer the run to a copied database.

During guarded transcript reads and writes, fully qualified keys such as
`agent:<agentId>:main` remain literal identities even when historical main-alias
metadata names another suffix. A queued turn rechecks its selected session and
lifecycle before restoring cold history and before appending, so a replaced
selection cannot restore or write the successor's history. These checks do not
rekey session rows or change database schemas.

Doctor preserves these qualified stored addresses and their parent, spawn, and
fork references when the configured main alias changes. It still repairs raw
aliases, delivery-proven legacy keys, and old default-agent main keys stored
under a replacement agent after the default agent was removed. Request aliases
continue to follow the current configuration.

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](https://funcoding.ai/agents/openclaw/cli/doctor/#session-sqlite-migration)
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.

Disk-backed history pages run their SQLite reads and display preparation in a dedicated session-transcript worker. Equivalent requests can share a queued read until worker execution starts; completed pages are not cached. The Gateway applies current profile display and rechecks session identity and access before publishing. Cold restoration and projection rebuilds remain with the existing Gateway storage owner. Incognito history stays in the Gateway process, and bound external CLI imports retain their local import owner. The HTTP history endpoint still returns the complete history when no limit is supplied.

The [proposed transcript working-set design](https://funcoding.ai/agents/openclaw/reference/session-management-compaction/resident-history/)
describes the remaining migration of live `SessionManager` history readers and
the consumer contracts required before enforcing resident budgets after writes.

## 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`