Does OpenClaw Remember Between Sessions? What Actually Persists

The short answer: No, not out of the box. Every OpenClaw session starts fresh, and nothing carries over unless the agent explicitly wrote it somewhere that gets loaded again. The built-in file-based memory is a setup project you configure and maintain, not a default. Scheduled cron jobs are the leakiest part: they often run without the memory files your chat session relies on.

Why does OpenClaw forget everything between sessions?

Sessions are isolated by design. When a session ends, the running transcript is gone, and the next session starts with a blank context window plus whatever files are configured to load at startup.

The built-in memory is best understood as a filing cabinet, not a brain. Markdown files get loaded into the prompt, the model reads them, and then the computation is gone. There is no learning step. If a fact was never written to a file, it does not become durable just because it came up in a previous conversation.

That is the first thing operators get wrong: the agent can handle a session brilliantly and still remember nothing the next day, because handling something well and saving it are two different actions, and only one of them happened.

What does OpenClaw's built-in memory actually do?

The community-settled setup is a three-tier file system:

1. MEMORY.md — a slim index at the workspace root, kept under about 3 KB, pointers only. Never full documents.
2. memory/ — a directory of topic notes (projects, people, decisions) that the agent searches with memory_search() for semantic lookup.
3. vault/ — deep storage for detailed notes, searched only when needed.

There is also memory_get() for exact reads of a specific note. With an embedding model configured (OpenAI, Gemini, Voyage, Mistral, or local Ollama), the search is semantic; without one, it falls back to keyword search.

All of this only works if the write side happens. memory_search can only find what was saved, and no component auto-saves by default. An agent that never writes is an agent with an empty filing cabinet and a very good search index over nothing.

Why do OpenClaw cron jobs forget even when chat remembers?

This is the question that bites automation operators specifically. Your interactive chat remembers the project layout, but the 7 AM cron job that triages your inbox acts like it has never met you.

Three reasons stack up:

Cron jobs do not reliably load the same context. The memory guide the community passes around calls this out directly: MEMORY.md is typically wired for the main session, while cron jobs and sub-agents should not be assumed to behave the same way. A workflow can feel forgetful even when the files themselves are fine, because the files were never loaded into that run.

Compaction wipes specifics. Around 75 to 86 percent of the context window, OpenClaw replaces the running transcript with a flat summary, and the model confidently misremembers details afterward. Long-running sessions rot from the inside.

There is a quiet bootstrap cap. Files loaded at session start get truncated around 20,000 characters, cut from the bottom, with no error anywhere. A MEMORY.md that grew past the cap silently loses its tail, which is usually where the newest notes live.

What are the options for giving OpenClaw persistent memory?

1. Run the MEMORY.md discipline. Keep the index slim and pointer-only, prune it weekly, and verify the agent actually writes after important turns. This fixes most forgetting and costs nothing. It breaks down the moment you stop curating, and the memory lives on one machine.

2. Set up the full three-tier system. MEMORY.md plus the memory/ directory plus vault/, with an embedding model for semantic search. This is the setup the optimization guides document step by step. It is the best free answer and the most maintenance.

3. Use the memos plugin. The open-source memos local plugin auto-captures every agent turn into a local SQLite database with hybrid keyword-plus-vector search and deduplication. It removes the manual-write problem. It is still local to the machine OpenClaw runs on.

4. Put the memory in a shared layer outside OpenClaw. This is the option for when the agent is not the only thing that needs to remember. Vilix AI is a cloud-hosted shared memory and work-state layer: OpenClaw connects as a headless agent with an API key, save_turn stores the full exchanges (not just distilled facts), and get_context loads the relevant context into the next run with semantic plus keyword retrieval. Because it is hosted, you manage no infrastructure, and the same memory follows you into Claude Code, Codex, and Cursor, so the cron job on your VPS and the chat on your laptop stop being strangers.

When should you look beyond OpenClaw's built-in memory?

Stay with the built-in files if one OpenClaw instance on one machine is your whole setup and you are willing to curate. Look beyond it when the memory needs to outlive the machine: OpenClaw on a VPS plus Claude Code on a laptop, a cron job whose findings a second agent has to act on, or full conversation history you can revisit instead of notes someone had to remember to write.

The failure mode to watch for is the quiet one. OpenClaw does not error when memory is missing or truncated; it just acts slightly off, confidently. If your cron jobs feel dumber than your chat sessions, check what each session type actually loads before you blame the model.

Frequently asked questions

Does OpenClaw have memory by default?

No. It ships with the machinery for file-based memory (MEMORY.md, memory_search, memory_get) but nothing is saved or loaded until you configure it, and the agent must explicitly write for anything to persist.

Can OpenClaw cron jobs access MEMORY.md?

Not reliably by default. Cron jobs and sub-agents run in sessions that should not be assumed to load the same startup files as your main chat. Verify what each session type loads; do not assume.

What is the difference between MEMORY.md and memory_search?

MEMORY.md is a small curated index that loads at session start. memory_search() is semantic search over the larger memory/ directory, used during a session when the agent needs something specific. One is a table of contents, the other is the library catalog.

Is there a plugin that gives OpenClaw automatic memory?

Yes. The open-source memos plugin for OpenClaw auto-captures every turn into a local SQLite store with hybrid search, removing the need for manual writes. It keeps everything on the local machine.