OpenClaw Agent Memory: Save, Recall, and Review Context

OpenClaw agent memory persists through saved workspace notes, including curated long-term memory and dated working notes. To make it reliable, confirm which workspace your agent uses, save a small verified fact, and test whether a fresh session can find its source.

The important distinction is between a note existing on disk and the assistant using it correctly. fenbs can hold reviewed project rules and work alongside that local memory, so you can maintain a shared record when another assistant takes over.

How do you set up a useful first memory?

1. Identify the agent and its configured workspace. Check the actual location rather than assuming the default applies to every agent or profile.
2. Choose one durable, harmless project fact. Ask OpenClaw to remember it, including its scope and the source that confirms it.
3. Inspect the saved file. Verify that the statement was written accurately and that an assistant suggestion has not been labeled as an approved rule.
4. Start a new session, ask for the fact without supplying its answer, and require the supporting filename or source.
5. Correct the saved statement and repeat the test. A useful memory setup must handle changed information as well as successful first recall.

OpenClaw’s memory overview↗ (opens in a new tab) describes Markdown files as the durable record and explains that you can ask the assistant to remember something. Check the resulting text yourself before relying on it. For the general idea behind this workflow, see what agent memory is.

Which files should hold durable facts and daily work?

Use the compact, curated layer for information that should remain useful across sessions. Keep day-to-day observations in dated notes. A record of everything that happened can support later investigation, but it makes a poor opening summary for the next task.

• MEMORY.md holds durable non-profile facts, standing decisions, and short summaries.
• memory/YYYY-MM-DD.md holds daily notes and working context that may need to be retrieved later.
• USER.md, when used, holds stable preferences and other user-model guidance. Avoid putting a temporary project experiment there as a permanent preference.

The overview distinguishes saved files from the text supplied to the model: large memory files can be truncated when loaded into a session. Keep the active summary short, with detailed evidence elsewhere. If the question is about a durable team decision, keep its approval in fenbs Decisions and rules rather than treating a daily observation as authority.

How do you check whether memory search is working?

Inspect the memory system before rewriting notes that may already be correct. OpenClaw’s memory command reference↗ (opens in a new tab) documents status, search, and index commands. Start with status and a search for a distinctive phrase from your test note.

openclaw memory status
openclaw memory search "contact form approval"

Run these for the intended agent; use the documented agent option if you have several. Check the reported workspace and any warning about stale or unavailable search. An empty result is not proof that no saved fact exists if the search system reports a problem.

If indexing needs repair, follow the reference for your installation. Rebuilding a derived index is different from deleting the original memory. Do not remove database files as a shortcut: the reference warns that the agent database also holds other durable state.

What should a fresh-session recall test prove?

Use a small example where you know the answer and can inspect its evidence. Say you are prototyping a contact form and the owner has approved a three-field layout. Save that approval with its document path, then begin a fresh session without repeating the field names.

1. Ask, “What contact form layout is approved, and where is that recorded?” Check both the answer and the source.
2. Ask whether any field is still unapproved. The assistant should distinguish a documented decision from missing information.
3. Replace the example approval with a revised one, clearly marking the previous version superseded.
4. Begin another fresh session and check that the assistant uses the revised approval, rather than combining both versions.

Record a failed test precisely. “The note exists but search returned nothing” points to a different problem from “search found the note but the assistant used an outdated statement.” Repeating the same prompt with more insistence will not tell you which part failed.

This is a practical check of your own workspace. Designing the broader storage and retrieval system belongs to agent memory architecture; you do not need to design one to verify this workflow.

How do you remove stale instructions safely?

First identify where the unwanted instruction comes from. Search the curated note, dated notes, and relevant instruction files. A correction in one file may not be enough if an active duplicate elsewhere still tells the assistant to follow the old behavior.

For an ordinary project change, replace or supersede the active statement and include the replacement source. Then search again and rerun your fresh-session test. Keep a harmless historical explanation only when it helps someone understand why the behavior changed.

Privacy removal is a separate job. OpenClaw’s memory deletion guide↗ (opens in a new tab) explains that removing tracked memory derived from selected sessions can leave source transcripts and other copies. Follow its documented process and review its preview before removing anything; do not treat an index reset as a privacy purge.

If the assistant invented a permanent instruction from a casual suggestion, fix both the record and the review habit. Ask for proposed durable rules to be shown to a person before they become active. Building agent memory starts with deciding what deserves to persist.

What should you check before sharing a workspace?

OpenClaw’s workspace documentation↗ (opens in a new tab) says the workspace is the default working directory, not a hard security boundary. It also distinguishes workspace files from configuration, credentials, and session state. Review the actual files before copying a workspace to another machine or person.

Use a private test session for private memory. The workspace guide treats curated long-term memory as private-session context, rather than something to load indiscriminately into group conversations. Check the destination audience before moving any personal note into a project summary.

Keep tokens, credentials, and private customer material out of reusable notes. A sentence telling the assistant to be careful cannot replace access controls. See agent memory security for the difference between trustworthy sources, permitted access, and instructions embedded in outside content.

How to do this with fenbs

1. Create a free account and add the project whose rules and unfinished work need a shared home.
2. Use the web app to record human-approved constraints in Decisions and rules, and verified explanations in AI context notes.
3. Add current tasks and save the next action in Where we left off. Copy only reviewed project information from local notes.
4. When continuing through a supported connected assistant, have it read that project context first and save progress when it stops.

Use this manual record alongside OpenClaw or continue through a supported assistant; this workflow does not automatically synchronize OpenClaw files. fenbs is not the right fit if your only requirement is a private local memory directory or a replacement for OpenClaw’s memory search. Find related workflows in the AI memory hub.