OpenClaw and embedding

[candid jetty]

Here's our previous conversation summarised by Krill

Brief summary:

Why you couldn’t see OpenClaw “thinking”
/think changes reasoning effort
/reasoning only shows reasoning if the backend actually returns it
/verbose is what shows tool/activity traces
with your local llama-server/... setup, OpenClaw was not guaranteed to expose reasoning blocks the way you expected

What we found about memory/embeddings
your memory files were real and indexed into ~/.openclaw/memory/main.sqlite
the search DB definitely contained the target file and matching text
but your live embedding backend and the stored index got out of sync

Root cause
the existing DB had 4096-dim vectors
your llama-server embedding endpoint was returning 2560-dim vectors
that mismatch breaks search
Best fix
use a separate remote embedding server (your 8081 llama-server is fine)
choose the embedding model you actually want to keep
do one clean reindex against that exact setup

Going forward
don’t keep full reindexing for normal use
“teaching” OpenClaw usually just means adding/editing Markdown in:
MEMORY.md
memory/*.md
let incremental indexing handle routine updates

If this reindex finishes and exact phrase search still fails, then it’s likely a real OpenClaw bug

[solemn garden]

To help others find answers, you can mark your question as solved via Right click solution message -> Apps -> ✅ Mark Solution

[gritty vale]

Nice, that summary helps.

This is the next thing I’d do:

openclaw --version
openclaw memory status --deep

And if you have not already done the clean rebuild after locking the embedding endpoint/model:

openclaw memory index --force

Important bit: keep agents.defaults.memorySearch pointed at the same embedding provider/model/dimension setup for both indexing and querying. If that changes, the index can go weird again.