Clawdguy
OpenClaw guide

OpenClaw memory: keep it reviewable, durable, and useful

OpenClaw memory is not magical “the model remembers.” It is explicit Markdown stored in the agent workspace. That is good news: you can audit it, edit it, and design it like a real system instead of a prompt trick.

TL;DR (the simple memory model)

If you only remember one thing: in OpenClaw, the files are the source of truth. The agent “remembers” what is written to disk. So the job is to decide what gets written, where, and when.

  • Durable facts and preferences go in `MEMORY.md`
  • Daily running notes go in `memory/YYYY-MM-DD.md`
  • Index/search is provided by the active memory plugin
  • Compaction happens when context is big; use memory flush to persist before summarization

How OpenClaw memory works (in practice)

OpenClaw memory is plain Markdown in the agent workspace. The model does not keep a secret long-term state. It reads selected memory files at session start, and tools can index/search the memory depending on the active plugin.

If you want to disable memory tools completely, you can disable the memory plugin slot.

  • Daily memory: `memory/YYYY-MM-DD.md` (append-only log)
  • Long-term memory: `MEMORY.md` (curated, stable, human-reviewable)
  • Workspace root is controlled by `agents.defaults.workspace`
  • Disable memory plugin with `plugins.slots.memory = "none"`

Default memory files

agents.defaults.workspace/
  MEMORY.md
  memory/
    2026-03-08.md

Disable memory plugin

{
  "plugins": { "slots": { "memory": "none" } }
}

What to store (and what to keep out)

Good memory improves future decisions. Bad memory pollutes prompts and creates contradictions. Keep long-term memory small, structured, and intentionally curated.

Good default

Promote notes into `MEMORY.md` only after they have proven useful more than once.

  • Store: durable preferences, decisions, operating rules, stable facts
  • Avoid: transient chat fragments, time-sensitive facts, raw web dumps, secrets without retention policy

Compaction and “memory flush” (prevent losing durable context)

Long-running sessions accumulate tool outputs and messages until the model context window gets tight. OpenClaw compacts older history into a summary entry, preserving recent messages. Before compaction, it can trigger a silent “memory flush” turn that reminds the model to write durable notes to disk.

This is the clean way to ensure the important parts of a long session become durable memory instead of getting buried in a summary.

  • Manual compaction: use `/compact` when a session feels stale
  • Auto-compaction: happens automatically near context limits
  • Memory flush: a silent reminder to write to `memory/YYYY-MM-DD.md` before compaction

Enable and customize memory flush

{
  "agents": {
    "defaults": {
      "compaction": {
        "reserveTokensFloor": 20000,
        "memoryFlush": {
          "enabled": true,
          "softThresholdTokens": 4000,
          "systemPrompt": "Session nearing compaction. Store durable memories now.",
          "prompt": "Write any lasting notes to memory/YYYY-MM-DD.md; reply with NO_REPLY if nothing to store."
        }
      }
    }
  }
}

How Clawdguy helps

When memory becomes part of production workflows, reliability depends on stable infrastructure and operational visibility. Clawdguy gives you a managed OpenClaw environment where you can run always-on workflows, inspect behavior, and evolve memory rules deliberately.

That is a faster path than stitching together local setups and hoping the context stays coherent over time.

  • Dedicated infrastructure for persistent workflows
  • Operational controls for logs, updates, and lifecycle management
  • A cleaner path from prototype memory to production memory