Memory and Proactivity: What OpenClaw Carries Forward

OpenClaw’s memory and automation systems answer two different questions: what from a conversation should remain useful later, and what should happen when a future time or event arrives. This chapter follows both lifecycles, from an observation in an agent turn to searchable or curated memory, and from a reminder request to scheduled work and delivery. The distinction helps you choose the right persistence mechanism without turning every durable fact into an instruction or every intention into a vague note.

The difference is important because an assistant’s continuity is shaped more by what it writes and triggers than by what it can retrieve in a single exchange. A system that automatically re-injects every past sentence may remember too much of the wrong kind. A system that stores “remind me when this happens” as ordinary prose may forget to act or act repeatedly. OpenClaw addresses those problems with separate tiers, provenance metadata, deterministic gates, and Gateway-owned lifecycle records.

Memory is a gated write path

OpenClaw’s default memory model begins with a simple rule: there is no hidden state. The assistant remembers what is written to inspectable workspace files and indexed in the memory system. That makes memory concrete, but it does not make all files equivalent. Instructions, durable facts, daily observations, future intentions, and review reports have different purposes and different paths into the model’s context.

The curated core contains compact durable material such as MEMORY.md and USER.md. These files may be loaded at the start of a session when their provenance is eligible and their size fits the bootstrap budget. Daily notes and session transcripts form an episodic tier: they can be detailed and append-friendly, but they are searchable on demand rather than injected into every ordinary turn. Standing intents and scheduled jobs are prospective state, because they describe conditions for future action rather than facts about the past. Dream reports are review artifacts for people; they are not supposed to become a second source of automatic promotion.

This separation solves a problem that a single “memory file” cannot. A daily note can record that an agent saw a web page, summarized a meeting, or tried a deployment. It is useful evidence, but it is not automatically a trusted instruction for tomorrow. A preference in the user model should be expressed as a current directive rather than as a long chain of contradictory historical observations. A timed reminder needs a scheduler, not a paragraph the model must remember to reread. OpenClaw gives these different jobs different owners.

The path from a turn to durable memory is deliberately staged. During ordinary work, the agent can write observations to a daily note. Before a long conversation is compacted, a memory-flush turn can save important context that has not yet been written, so summarization does not silently discard it. When a session ends, eligible transcript material can enter the episodic corpus. None of these steps means that a fact has become part of the curated core. It has entered a pool from which a later consolidation process may draw.

That later process is dreaming. Its light phase sorts and stages recent signals; its REM phase reflects on themes and recurring ideas; its deep phase ranks candidates and promotes only those that pass deterministic thresholds. Candidates are evaluated using signals such as relevance, repeated recall, query diversity, recency, recurrence across days, and conceptual richness. The model is used inside the bounded consolidation step to choose additions, merges, or supersessions against the existing curated file.