How to Check OpenClaw Memory
When an OpenClaw agent forgets something it should know, the fix starts with looking at what's actually in its memory...
When an OpenClaw agent forgets something it should know, the fix starts with looking at what's actually in its memory — and what's actually loaded into context. These are two different things, and the gap between them explains most "memory isn't sticking" reports.
Here are the commands to inspect both.
How do you check what OpenClaw remembers?
Note on commands: OpenClaw's CLI changes between versions. The commands below reflect a recent CLI, but flag names and subcommands shift — run
openclaw --help(andopenclaw memory --help) to confirm what's current in your install before relying on any specific command.
Run openclaw memory status. It reports the health of your memory plugin and index.
For a deeper inspection:
openclaw memory status --deep— probes the vector store and embedding-provider readinessopenclaw memory status --fix— repairs stale recall locks and normalizes promotion metadataopenclaw memory status --json— machine-readable outputopenclaw memory status --agent main— scope to a specific agent
If status shows the store is dirty, openclaw memory status --index checks and automatically reindexes. To force a full rebuild from source documents: openclaw memory index --force.
How do you see what's loaded into context?
This is the command most people miss, and it diagnoses the majority of memory problems: run /context list inside your OpenClaw session.
It shows every file currently loaded into the agent's working context — which memory files were read at session start, how much of each is present, and how close you are to the limits. /context detail goes deeper, and openclaw doctor shows raw versus injected sizes plus truncation status.
What to look for:
- Missing files. If
MEMORY.md,USER.md, orAGENTS.mdaren't listed, they're having zero effect on the agent. - Truncation. If a file shows "TRUNCATED," it exceeded the 20,000-character per-file cap and the agent only sees part of it.
- Raw vs. injected mismatch. If the raw and injected character counts differ for a file, content is being cut. This is the most common invisible problem — your file looks fine on disk, but the agent works with an incomplete copy.
How do you search OpenClaw's memory?
openclaw memory search "your query" runs a semantic search across your memory files. Or use the flag form for more control:
openclaw memory search --query "deployment notes" --max-results 20
This searches everything indexed — MEMORY.md, daily logs, the whole memory directory — by meaning rather than exact keywords. It's how you confirm whether something you expected the agent to know is actually stored and retrievable. If a search returns nothing, the information either was never saved or isn't indexed.
Why does memory status show less than expected?
Three usual causes, all worth checking in order:
The daily notes window. Only today's and yesterday's daily notes auto-load. Older notes exist on disk and are searchable, but they're not in context unless retrieved. Status showing little loaded context can simply mean the agent hasn't searched for older material.
Truncation. A file past its character cap is cut in the injected copy. Disk looks complete; context isn't.
A dirty or unbuilt index. If semantic search was recently enabled or files changed, the index may be stale. openclaw memory status --deep --index checks and rebuilds.
What if you're checking because memory keeps disappearing?
If you're running these commands often because the agent keeps losing track of who you are and what you're working on, the diagnostics are treating a symptom.
Foundational context — your role, your priorities, who you are — shouldn't depend on whether the agent wrote it to the right file and can search it back. That kind of context is better delivered deterministically, loaded at the start of every session regardless of what the agent remembered to save.
→ Why that's a different problem: AI Memory vs. AI Context: What's the Difference?
→ How to fix it at the source: How to Make OpenClaw Memory Better
