Skip to main content
Memory now lets an eligible personal Claw recall relevant context from that agent’s other private conversations, including what mattered immediately before a reset, while visible workflows let you search indexed sources, inspect how memory is working, import supported history, and remove attributable derived memory. Recall stays within the same agent’s private conversations and respects explicit isolation and access policy. Built-in Memory owns the core search and recall path, with a supported Doctor migration from QMD. LanceDB, Memory Wiki, external embedding services, MEMORY.md, and USER.md still have distinct roles, and forgetting derived memory does not erase the original conversation or copies outside OpenClaw.
Built-in Memory now owns the core search and recall path. If you use QMD, run openclaw doctor --fix to remove retired QMD settings, carry forward supported extra paths and any session indexing you explicitly enabled, preserve compatible rows already in the agent database, and rebuild the index from canonical Markdown.The migration carries supported data into a different core, so QMD-only reranking, query expansion, and cross-agent transcript search are retired. Malformed structures, incompatible vector dimensions, and data without a safe owner remain stopped for repair.
On eligible personal setups, your Claw can recall relevant context from that same agent’s other private conversations by default, including what mattered immediately before you reset the session. Recall remains limited to that agent’s private conversations; groups, channels, shared aliases, other agents, deleted history, and policy-blocked sources stay out, and an explicit direct-message isolation setting still wins.Built-in search now understands filenames, full and partial Unicode paths, and configured extra paths, broadens thin strict matches, and keeps keyword results available when an optional embedding provider cannot start. While memory or session content is being rebuilt, each search stays on one stable published index, and later searches use its replacement only after publication instead of waiting behind routine maintenance or mixing generations. Search stays within the configured roots and agent boundaries, while required embedding providers fail closed.Sessions without a configured reset policy now remain open across days, and durable reset or compaction markers explain visible history changes. SQLite-backed chats on web, macOS, iOS, and Android can rewind to a user message, fork the conversation, and switch among preserved branches. Rewinding changes the transcript branch, but it does not undo files, sent messages, or other tool side effects.
Importing is separate from the QMD-to-built-in upgrade. You can bring supported memory from Codex, Claude Code, or Hermes into an agent workspace through the Control UI, first-run setup, or Ask OpenClaw while leaving the source alone and not sweeping in credentials, settings, skills, or arbitrary provider files. If the destination already contains conflicting content, replacement has to be reviewed explicitly.Old conversations follow their own preview-first path. The CLI shows what it would stage before it writes, the Control UI reports bounded-batch progress, and material owned by that import can be rolled back and applied again. Large histories use bounded or indexed processing, Memory Wiki preserves human notes through supported imports, and older history without complete ownership tracking may need to be rescanned. Staged material still does not become durable memory until dreaming or an explicit promotion chooses it.
Settings now has a Memory destination for live status, Dreams, configuration, indexed-source search and browsing, add-on controls, and embedding readiness, so you can see what your Claw has available and open safe workspace memory files. Memory Wiki overview and Imported Insights dashboards now read published snapshots instead of reparsing the whole vault for every request, keeping large views responsive and showing when a rebuild is running or a manual compile is required. Session and legacy sources remain snippet-only, and add-on changes remain read-only unless the connection has admin authorization.Eligible observations can be consolidated into MEMORY.md, durable directives can live in USER.md, standing intents can wait for the right event, and project memories stay scoped to the project that produced them. Automatic memory remains bounded and provenance-gated, so content derived from network or restricted sessions keeps an untrusted origin and stays out of automatic context even though an explicit search can still surface it. This provenance protection applies to newly tracked material, while older untracked files retain their existing classification.openclaw memory forget lets you preview a purge by session, hook source, or participant, then remove attributable derived memory and stop the selected session from being pulled back in by backfill, indexing, or dreaming. An interrupted purge can be retried, and newly attributable consolidation or backfill diary claims are removed with the selected session. The purge follows recorded provenance, so the original transcript, older lineage-free notes, other agents’ stores, direct or external writes, exports, and backups may remain. Review the preview before applying it.