OpenClaw Learning Memory

Always-on memory for OpenClaw agents.

A context engine that injects a curated MEMORY.md + USER.md block into every agent turn — including group and channel sessions, which OpenClaw's default memory recall excludes. This is why bots in Zalo/Telegram groups "forget" context and rules over time; this plugin fixes that by loading your curated memory on every run (an always-on prompt-memory layer).

What it does

• Always-on: reads the agent's curated memory files and prepends them to the system prompt on every run (all sessions, groups included) via systemPromptAddition.
• Curated, not bloated: a tight character budget (default 3500) forces the agent to curate — a small sharp memory beats a huge one. Overflow is trimmed, not accumulated.
• Safe & thin: it does not reinvent history storage or compaction. History is passed through untouched and compaction stays with the runtime (ownsCompaction:false). Any read hiccup degrades to "no injection" for that turn; a hard error quarantines the engine and the host falls back to the legacy engine — it can't brick an agent.

It does not need a remote memory database. Your memory lives in local workspace files that the agent already writes (MEMORY.md, USER.md).

Install

clawhub package install tuanminhhole/openclaw-learning-memory

Then select it as the agent's context engine in openclaw.json:

{
  "plugins": { "entries": { "learning-memory": { "enabled": true } } },
  "agents": {
    "defaults": {
      "plugins": { "slots": { "contextEngine": "learning-memory" } }
    }
  }
}

Config

Under plugins.entries.learning-memory.config:

How the agent keeps memory good

The engine only loads memory; the agent curates it. Pair this with an agent instruction to proactively write durable facts and rules into MEMORY.md (short, distilled bullet points) and keep it tight. Because the file is now injected every turn, what the agent writes actually comes back — so it compounds instead of evaporating.

License

MIT © tuanminhhole