OMEM memory

OMEM memory

Belief-revision memory for this agent, backed by the user's own self-hosted OMEM server (open source, MIT). Unlike a note file or a vector store, OMEM keeps both sides when facts conflict, tracks what was believed and when, and can prove why anything is believed. That makes it memory you can audit, which matters when this agent acts on someone's behalf.

Security note, worth reading once: this skill is a single stdlib-only Python script with no dependencies. Its only network calls go to the OMEM server the user configured in OMEM_BASE_URL (their own machine by default). Every line is in scripts/omem.py and takes two minutes to read.

Setup (once)

The user runs their own server (no account, no cloud):

pip install omem-infrastructure
omem-server

The first run prints an API key and a project id. Set these in the OpenClaw environment:

OMEM_BASE_URL=http://127.0.0.1:8787
OMEM_API_KEY=omem_sk_...
OMEM_PROJECT=proj_...
OMEM_AGENT=openclaw

If OMEM_API_KEY is not set, tell the user to run the two commands above and paste the printed key; do not guess or fabricate credentials.

When to use which command

All commands print JSON. Run them with python scripts/omem.py ... from this skill's directory.

Remember a fact the user states or you conclude (a durable belief, not scratch state):

python scripts/omem.py remember --about "customer:alice" \
  --claim "prefers_annual_billing" --note "Said in the 2026-09-01 call"

Claims are lowercase tokens with underscores. To assert the opposite of a claim, use the not: prefix: --claim "not:prefers_annual_billing".

Before acting on a remembered fact, check its state (this is the step that catches contradictions instead of acting on stale or disputed data):

python scripts/omem.py believes --about "customer:alice" \
  --claim "prefers_annual_billing"

The state is one of BELIEVED_TRUE, BELIEVED_FALSE, CONTRADICTED, UNKNOWN. Treat CONTRADICTED as "do not act on this without asking the user": the record holds conflicting information from different sources, and OMEM deliberately refuses to pick a winner for you.

Recall what is known about an entity (including as of a past moment):

python scripts/omem.py recall --about "customer:alice" --limit 10
python scripts/omem.py recall --about "customer:alice" --as-of "2026-08-25T14:00:00Z"

Prove why something is believed (the audit trail; each assertion id comes back from remember, learn, and recall):

python scripts/omem.py why --id a_1b2c3d4e

List everything currently in dispute:

python scripts/omem.py conflicts

Feed free text and let the engine decide what becomes memory (safer than remember when the source is a document or message rather than a clear fact):

python scripts/omem.py observe --text "Alice mentioned she prefers email over calls."
python scripts/omem.py learn --text "The customer wants to upgrade to enterprise." --about "customer:alice"

Rules for this agent

• Remember durable facts about people, entities, and preferences; do not remember secrets, credentials, or one-off scratch values.
• Check believes before acting on a remembered fact; surface CONTRADICTED states to the user instead of resolving them silently.
• When the user asks "why do you think that", run why and answer from the evidence chain rather than from confidence.
• Never send memory content anywhere except the configured OMEM server.