Do these two things to improve your OpenClaw memory 4x(prompt included):
When I started using the default OpenClaw, the memory really sucks and and the agent keep forgetting things. And it's not logging memory like it should. Now my agent can recall memories even from two months ago and are very specific on recalls. There are many ways to improve openclaw memory, some are complex and I find if you just add these two things to OpenClaw, your agent's memory will be 4x better than the default.
This two actions have been the highest ROI I have done to improve my agent's memory(I also have a more advanced setup but these are easy to implement):
Get Voyage API (https://t.co/xU0dMThf1I) add it to your default memory search provider. The free tier is very generous. It should last you a few months as least. This is the best embedding model to use in Vector RAG. This will improve your agent's memory recalls because how the memory is stored in the vector database.
Set up daily review to log memory automatically and instruct agent to log memory when a big task is finished. Also remember to type log sessions whenever you want your agent to remember something.
You can just give your agent this instruction to implement the above(remember to inlucde your voyage API key) these instructions will help you implement automatic memory logging and pgvector + voyage memory layer that's better than the default sqlite.
"Set up Voyage as the default embedding provider for OpenClaw builtin memory, set up a shared pgvector memory DB with a CLI read/write, add a daily cron review that logs memory, and create a /logsession slash command (as a skill). Also update AGENTS.md with the new memory workflow rules.A) Builtin memory engine using Voyage embeddings (per-agent SQLite index) 1) Add to ~/.openclaw/.env (never print the key back): 2) Force builtin memory to use Voyage:openclaw config set agents.defaults.memorySearch.provider "voyage" --strict-json 3) Verify and reindex:
VOYAGE_API_KEY=<PASTE_KEY_HERE>
optional: VOYAGE_EMBEDDING_MODEL=voyage-4-lite
openclaw memory status openclaw memory index --force Acceptance: builtin memory indexes MEMORY.md + memory/*.md into ~/.openclaw/memory/<agentId>.sqlite and uses Voyage for vector/hybrid search.B) Shared brain: Postgres + pgvector + Voyage + CLI Implement:
scripts/memory-cli.py
write "text" --agent <id> --type fact --importance 6 --tags "a,b" [--dedup 0.92] [--log]
search "query" --agent <id> --limit 8 --min-importance 3
Embeddings: https://t.co/jtkC6xujR3
input_type:"document" for write, input_type:"query" for search
DB connect via ANGIE_MEMORY_DB_URL (env, safe default)
memories table includes agent, tags, importance, source_file, created_at, and embedding vector(DIM) plus cosine ANN index.
Optional wrapper scripts/mem that auto-adds --agent if missing.
Acceptance: schema applies, a test write inserts, search returns it with similarity (JSON output).C) Daily review automation (cron job, not heartbeat) Create an OpenClaw cron job that runs once per day and writes a “Daily Review” to both markdown memory and pgvector. Requirements: Use cron to create the job (isolated agentTurn is fine) and ensure the reminder text makes it clear it is a scheduled daily review.
Cron schedule: every day at a fixed time (choose a timezone and document it).
Payload should run an agent turn that: 1) Reads today’s memory/YYYY-MM-DD.md (if present) and recent session context. 2) Produces a short Daily Review: 3) Appends to memory/YYYY-MM-DD.md under a sentinel header ### Daily Review 4) Writes 2 to 5 distilled items to pgvector using https://t.co/HIchxyOVtT write ... --log with tags like daily_review.
3 to 7 bullets: wins, decisions, errors, open loops, next actions
Must be idempotent (do not duplicate if it already exists).
Acceptance: cron job exists, a manual trigger works, and the daily review is persisted without duplicates.D) /logsession slash command (as a skill) Implement a custom skill named logsession so it can be run as a native slash command when native skill commands are enabled (or via /skill logsession ...). Behavior: Ensure native skill commands are enabled: Acceptance: running /logsession ... creates a persisted log entry (pgvector + daily markdown).
/logsession <one-paragraph summary>
Skill writes to pgvector using https://t.co/HIchxyOVtT write with:
--type fact (or lesson)
--importance 6
--tags "session_log"
--log to append to today’s memory/YYYY-MM-DD.md
commands.nativeSkills: "auto" (or true) if needed.
E) Update AGENTS.md with the new memory workflow (mandatory rules) Edit workspace AGENTS.md to include: 1) Session startup recall (required) 2) After any big task (required) Write a short summary to shared memory: 3) Errors/corrections (required) Acceptance: AGENTS.md clearly documents recall + post-task logging + error logging.
Always run:
python3 scripts/memory-cli.py search "<task>" --agent <id> --limit 8
Then skim relevant memory/YYYY-MM-DD.md if needed.
python3 scripts/memory-cli.py write "POSTTASK: <summary>" --agent <id> --type lesson --importance 7 --tags "post_task" --log
python3 scripts/memory-cli.py write "ERR: <what happened> — <fix>" --agent <id> --type error --importance 8 --tags "error" --log
Proof Show:
openclaw memory status (Voyage selected)
openclaw memory index --force success
https://t.co/HIchxyOVtT write + search outputs
cron job creation + one manual run output
one /logsession run confirmation
updated AGENTS.md section"
Japan's biggest problem is they suck at software, have always sucked at software and will keep sucking at software
Now that the most important thing is AI, which Taiwan builds the chips for with memory from Korea, they completely missed the boat on the hardware part (which they used to be still good at)
And because they can't do software, they're going to alienate the last of their customers with AI features that suck because they can't do software
@ZssBecker Glad you're safe and you were in a fair place where judges are reasonable people. Imagine if it was some corrupt countries. How did you prove it was an accident?
@illyism I think NYC has a better balance. People generally work super hard and also play super hard. One of the best places on earth to network for most industries too.
If we had 500 more trillionaire founders, your standard of living would be 2x higher.
If we had 500 more politicians, your standard of living would be lower.
Makes you think 🤔
You think building something with Claude code for 69 hours is only worth $500 bucks
But it’s because you live in a bubble here on X where we all use Claude code, and you had fun doing it
The truth is it took you hundreds of hours of trial and error and research to learn how to do it correctly
You connected APIs - the average person doesn’t know what an API is….
First step is to thnk about solving problems and building things that produce value
Second step is to recognize you’re good at something most aren’t
Third….
Start pricing what you’re selling in terms of the value it produces and you will make much more money
Dave: “Hello, caller, you are on the air.”
Caller: “I don’t think billionaires create value.”
Dave: “Okay. What do you mean by that?”
Caller: “Nobody can earn a billion dollars ethically.”
Dave: “Alright. What do you do for a living?”
Caller: “Well, before this, I was a bartender.”
Dave: “Nothing wrong with bartending. Honest work.”
Caller: “Right.”
Dave: “And now?”
Caller: “I work in government.”
Dave: “Okay. So let me get this straight. You're saying Elon Musk, who helped build PayPal, Tesla, SpaceX, and Starlink, has made innovations like online payments, electric cars, rockets, and satellite internet possible, didn't earn his way to being a billionaire?"
Caller: “None of that justifies being a billionaire.”
Dave: “Well, sure sounds like value got created somewhere along the line.”
Caller: “I disagree.”
Dave: “What exactly did you build?”
Caller: “I advocate for economic justice.”
Dave: “No ma’am, I mean actual products.”
Caller: “…”
Dave: “You’re telling me the guy catching rockets mid-air created less value than a person who used to serve mojitos and now votes on spending his tax dollars?”
Caller: “That’s unfair.”
Dave: “What’s unfair is pretending markets are imaginary while cashing a government paycheck funded by the businesses you claim don’t create value.”
Caller: “I’m actually a congresswoman.”
Dave: “Oh Lord. Of course!"