Cognimemo

Appearance

Python SDK ​

cognimemo-client — the official Python client.

bash
pip install cognimemo-client

Quickstart ​

python
from cognimemo_client import Cognimemo

cm = Cognimemo(base_url="https://api.cognimemo.com", api_key="cmk_live_…")

cm.retain(bank_id="[email protected]", content="Jane ships on Fridays.")
res = cm.recall(bank_id="[email protected]", query="when does Jane ship?")
for r in res.results:
    print(r.type, r.text, r.occurred_start, r.scores)

Prefer await cm.aretain(...) / await cm.arecall(...) in async code — every method has an a-prefixed async version. For clean shutdown, use it as a context manager:

python
with Cognimemo(base_url="https://api.cognimemo.com", api_key="cmk_live_…") as cm:
    cm.retain("[email protected]", "…")

Data model — space → bank → memories ​

python
cm.create_space("acme", name="Acme Corp")
cm.retain(bank_id="[email protected]", content="Jane prefers dark mode", space="acme")
cm.recall(bank_id="[email protected]", query="what does jane prefer?", space="acme")

cm.list_spaces()
cm.list_space_banks("acme")
cm.add_bank_to_space("acme", "[email protected]")
cm.remove_bank_from_space("acme", "[email protected]")   # person + memories kept

Typed memory blocks ​

Each helper stores a distinct memory type that's processed and recalled differently — procedure (how a task is done), reasoning (why a decision was made), preference (how the user wants things), correction (a "no, do it this way" signal), profile (the user's tools/access). See Memory types & typed blocks for what each means and when to use it.

python
cm.retain_preference("[email protected]", "For day-planning, use personal Gmail, never work mail.")
cm.retain_procedure("[email protected]", ["make build", "kubectl apply -f prod.yaml"], rationale="prod deploy")
cm.retain_reasoning("[email protected]", "Chose gRPC over REST: p99 latency mattered more.")
cm.retain_correction("[email protected]", "No — drain the pod before restart.")
cm.update_profile("[email protected]", "Admin on AWS, GitHub, Grafana.")

# any type directly:
cm.retain(bank_id="[email protected]", content="Acme HQ is in Berlin.", fact_type="world")

# type-filtered recall:
cm.recall(bank_id="[email protected]", query="how to deploy", types=["procedure"])

See Typed memory blocks for the full list.

Layered / org memory ​

python
cm.retain(bank_id="__org__:acme", content="Deploy freeze every December.")
cm.recall(bank_id="[email protected]", query="deploy policy", space="acme", include_org=True)
# or:
cm.recall_layered("[email protected]", "deploy policy", space="acme")

Entities ​

python
cm.retain(bank_id="[email protected]",
    content="Jane joined the DeepMind team on Gemini.",
    entities=[{"text": "Jane", "type": "PERSON"}, {"text": "Gemini", "type": "PROJECT"}])

cm.entities.list_entities("[email protected]")

# no-LLM auto-entities for verbatim ingest (off by default):
cm.update_bank_config("[email protected]", retain_auto_entities=True)

Encryption at rest ​

Encryption is a project-level policy set in the console (Settings → Security), not an SDK call. When it's on, memories are stored encrypted and the SDK keeps returning plaintext — nothing changes in your code. See Encryption at rest.

Recall response ​

recall(...) returns RecallResponse with .results (a list of MemoryFact) and .usage. Each MemoryFact:

fieldmeaning
typefact type (world … profile)
text, contextthe memory body (decrypted transparently)
occurred_start / occurred_end / mentioned_attemporal anchors
entitieslinked entity names
scores{final, semantic, keyword}
metadata, tags, document_id, chunk_idprovenance

.usage reports {tokens_used, max_tokens, truncated} (and org_blended for layered recall).

Reflect ​

python
answer = cm.reflect(bank_id="[email protected]", query="What is Jane's release cadence?")