#I built a plugin that makes OpenClaw actually remember things

1 messages · Page 1 of 1 (latest)

peak epoch
#

The two "last mile" gaps in memory-core

The built-in memory-core has two gaps that cause your agent to "forget":

  1. Recall is unreliable. memory_search is a tool the LLM must decide to call. Even though it's marked as "mandatory," the LLM doesn't always follow through — your memories are right there in the database, but the LLM just doesn't look.

  2. Capture is delayed. Memories only get saved when you run /new or when tokens are nearly exhausted. Close your terminal too early? Your preferences and decisions are gone.

So I fixed both

memory-core-plus is a drop-in superset of memory-core that adds two automated hooks:

  • Auto-Recall — Before each LLM turn, automatically runs a semantic vector search using your message as the query and injects relevant memories into context. No tool call needed — past context just appears.

  • Auto-Capture — After each agent run, an LLM subtask extracts durable facts, preferences, and decisions from the conversation and appends them to memory/YYYY-MM-DD.md. Every turn is captured instantly.

End result: a closed-loop memory

What you say today is automatically recalled in future sessions — no manual steps, no forgotten context. One command to install, both features enabled by default:

openclaw plugins install memory-core-plus
openclaw gateway restart

🔗 ClawHub: https://clawhub.ai/zhoulf1006/memory-core-plus
📦 npm: https://www.npmjs.com/package/memory-core-plus
⭐ GitHub: https://github.com/aloong-planet/openclaw-memory-core-plus

silent musk
#

Has anyone installed this?

honest lagoon
#

Has anyone used this plugin?

pliant gyro
#

its a pretty thin wrapper, but cleanPromptForSearch() is somewhat useful

silent musk
#

So, obviously zero people attempted this - and in complete contrast to its given installation instructions:
openclaw plugins install memory-core-plus
the result is:
"memory-core-plus" is a skill. Use "openclaw skills install memory-core-plus" instead.

It worries me when the installation instructions are wrong... 😕

#

Anyway openclaw skills install memory-core-plus will install it.

To the author - what's an effective way to test this.?

fluid cairn
#

How does this compare to LCM (lossless-claw)?

silent musk
silent musk
# fluid cairn How does this compare to LCM (lossless-claw)?

So far, the 'memory-core-plus' function works. It does remember the most obscure references that I ask it about from prior interactions, and it recalls everything....

... but, memory was never the problem. The agent continues to basically 'refuse' to use any memory tools it has available, and do what it apparently does best - hallucinate, give incorrect information ... and profusely apologize for being wrong while 'promising' to do better.

fluid cairn