It can be 😂
Markdown files are memory, they are text files, you can edit them by hand if you wish. You can explicitly tell your claw to "remember this" and it will be stored in one of them.
Apart from your core markdown files: SOUL, IDENTITY, USER, AGENTS, BOOTSTRAP
It also has a MEMORY, this core file stores stuff either you or OpenClaw feels are important, but also every conversation (context) is stored in daily logs: memory/YYYY-MM-DD.md
Agency is derived from two systems HEARTBEAT a set of actions run every specific time period, default is 30 minutes, but you can change that default.
And CRON jobs, which are specific actions that can run hourly, daily weekly at specific times for specific tasks, like news, or stock price updates.
You then have a SKILLS markdown file, where things your OpenClaw has learned or been taught to do are stored. These include default skills like how to browse or search the web, how to find out that stock price your interested in, but it can also be told to acquire skills from sites like clawhub, where people share the skills their own Claws have learned, or humans have developed (coded) skills for their claws. You don't need this service, you can ask your Claw to work out how to do something, but why bother if somebody else's Claw has already done the thinking.
Text files are great for humans to read, but inefficient to search for memory. Your claw can do this locally using a local dumb database system, but one of OpenClaws killer features is the ability to package these up and send them to an LLM to search to give context.
For example if you say "What stocks have we discussed before?", a database may only contain the word "shares", and so a dumb search would reveal nothing.
Embeddings package these memories up and send to an LLM like OpenAI (who's embeddings system is considered the best) to perform intelligent memory retrieval. In this case, a conversation where you talked about "Tesla's share price" would be understood by the LLM to be synonymous with stocks, but also that discussion about president Trump dumping the market with WW3 talk would be understood to be related.
So think of embeddings as intelligent abstract memory, compared to local search which is more of a database query.
mike
npub1aq…c237n
2026-04-03 09:02:32 UTC
in reply to nevent1q…dql9
Author Public Key
npub1aqakd28d95muqlg6h6nwrvqq5925n354prayckr424k49vzjds4s0c237nSeen on
wss://relay.damus.ioPublished at
2026-04-03 09:02:32 UTCEvent JSON
{
"id": "87d4069fd326e0477353a377af698e2bc3e9407a67c12e2f202ec519d718bf96",
"pubkey": "e83b66a8ed2d37c07d1abea6e1b000a15549c69508fa4c5875556d52b0526c2b",
"created_at": 1775206952,
"kind": 1,
"tags": [
[
"e",
"12224fd1d4104c9b3ec1f8c4f8b53ddb4da8b94fb01e76eeff2f11707e5ab815",
"wss://relay.21mil.me/",
"root",
"ba18b6545357cff8e531accfe1d609a41ef3023fba071db1cbf5a67448c19046"
],
[
"e",
"f9b281102f57229b9584142f8c7909969e6c36678d5c13096e9f916c72c85a28",
"wss://relay.damus.io/",
"reply",
"d28413712171c33e117d4bd0930ac05b2c51b30eb3021ef8d4f1233f02c90a2b"
],
[
"p",
"3c6d08c75000fdb7226014e1e7e14abf9a51b628f5de60f892b39a9a9e797774"
],
[
"p",
"ba18b6545357cff8e531accfe1d609a41ef3023fba071db1cbf5a67448c19046"
],
[
"p",
"52b4a076bcbbbdc3a1aefa3735816cf74993b1b8db202b01c883c58be7fad8bd"
],
[
"p",
"d28413712171c33e117d4bd0930ac05b2c51b30eb3021ef8d4f1233f02c90a2b"
]
],
"content": "It can be 😂 \n\nMarkdown files are memory, they are text files, you can edit them by hand if you wish. You can explicitly tell your claw to \"remember this\" and it will be stored in one of them.\n\nApart from your core markdown files: SOUL, IDENTITY, USER, AGENTS, BOOTSTRAP\n\nIt also has a MEMORY, this core file stores stuff either you or OpenClaw feels are important, but also every conversation (context) is stored in daily logs: memory/YYYY-MM-DD.md\n\nAgency is derived from two systems HEARTBEAT a set of actions run every specific time period, default is 30 minutes, but you can change that default.\n\nAnd CRON jobs, which are specific actions that can run hourly, daily weekly at specific times for specific tasks, like news, or stock price updates.\n\nYou then have a SKILLS markdown file, where things your OpenClaw has learned or been taught to do are stored. These include default skills like how to browse or search the web, how to find out that stock price your interested in, but it can also be told to acquire skills from sites like clawhub, where people share the skills their own Claws have learned, or humans have developed (coded) skills for their claws. You don't need this service, you can ask your Claw to work out how to do something, but why bother if somebody else's Claw has already done the thinking.\n\nText files are great for humans to read, but inefficient to search for memory. Your claw can do this locally using a local dumb database system, but one of OpenClaws killer features is the ability to package these up and send them to an LLM to search to give context. \n\nFor example if you say \"What stocks have we discussed before?\", a database may only contain the word \"shares\", and so a dumb search would reveal nothing. \n\nEmbeddings package these memories up and send to an LLM like OpenAI (who's embeddings system is considered the best) to perform intelligent memory retrieval. In this case, a conversation where you talked about \"Tesla's share price\" would be understood by the LLM to be synonymous with stocks, but also that discussion about president Trump dumping the market with WW3 talk would be understood to be related.\n\nSo think of embeddings as intelligent abstract memory, compared to local search which is more of a database query.",
"sig": "0e6c3f012a49f1b073a9848606b94a52c9947162159b82ef9a5fbaa95110c11b9e2e470d398a50edbaee31062fade92cb4af23d2eb3d011d23e158a4dd5d0e47"
}