guidesJune 17, 2026· 5 min read

Giving an agent memory from Python: CMEM in practice

The memory commands from application code — store facts, recall them semantically, and watch consolidation retire the stale ones. A worked example in Python.

Agent memory is a few commands, and from Python it's a few method calls. The pattern: extract durable facts from a conversation, store them scoped to (agent, user), and recall them by meaning on the next turn — letting consolidation keep the picture current.

store, consolidate, recall

from crowkis import CrowkisClient

mem = CrowkisClient(host="127.0.0.1", port=6379)
AGENT, USER = "support", "u_42"

# learn three things across a conversation
mem.cmemset(AGENT, USER, "prefers email over phone")
mem.cmemset(AGENT, USER, "moved to Berlin in March")
mem.cmemset(AGENT, USER, "no longer in Munich")   # retires the old location

# recall by meaning, top-3, recency-blended
facts = mem.cmemget(AGENT, USER, "where does this customer live?", k=3)
print(facts[0])   # -> "moved to Berlin in March"

Because memory consolidates, you don't have to hunt down and delete the stale fact — storing the contradicting one retires it automatically. The recall is ranked by relevance blended with recency, so the current answer surfaces first.

extract facts, and honour erasure

# pull durable facts straight out of a transcript (deterministic, no model call)
mem.cmemextract(AGENT, USER, transcript_text)

# bi-temporal: what did we believe on April 1st?
mem.cmemasof(AGENT, USER, "address", at="2026-04-01")

# right to be forgotten
mem.cmemforget(AGENT, USER, "payment details")

In plain words: Tell the agent facts, ask by meaning, and let it retire what changed. Storing 'moved to Berlin' quietly forgets 'lives in Munich' — no manual cleanup.