Session observers
Each run receives a zero-semantic receipt with health tier, exact-cycle evidence, work conversion, review boundary, and boundary flags.
CAUM Live Workflow Memory
Agent workflows need memory, not just traces.
CAUM observes repeated agent sessions as one customer-scoped workflow. Small observers review each run; a larger workflow observer remembers recurrence, task-family scope, conversion drift, durable fingerprints, and long-horizon rhythm without reading private content.
Structural memory
Zero-semantic
Observe-only
No blocking
No truth judgment
The product shift
A receipt tells you what happened in one run. Workflow Memory tells you whether the same structural pressure is returning, worsening, or forming a larger rhythm across repeated work.
Workflow memory
Repeated sessions feed a customer-scoped memory layer that tracks durable fingerprints, cross-session echoes, task-family scope, and conversion drift.
Review cards
The dashboard prioritizes workflows that deserve human review first, without claiming answer correctness or authorizing automatic control.
Controlled workflow-memory demo
These are controlled engineering probes, not customer rates or ROI proof. They show how the product behaves across useful workflow families while keeping CAUM inside the Truth Lock.
Healthy repeated workflow
QuietEight repeated healthy sessions stayed T1 with no hard alerts and no workflow review card.
Clean benign recurrence
QuietA productive repeated workflow preserved its durable fingerprint without escalating the workflow card.
Exact retry loop
High reviewRepeated exact retry pressure surfaced at session level and workflow level with a high review card.
Noisy diluted loop
High reviewInjected noise did not hide this controlled workflow family from the larger observer.
Gradual degradation
Drift watchThe first half stayed quiet; the second half became reviewable as work conversion degraded across sessions.
Long-horizon rhythm
Long watchThe temporal swarm surfaced a larger rhythm across sessions, not only short loops inside one run.
Signals that make memory useful
The most useful signals are not only exact cycles. Workflow Memory watches whether structural conversion changes over time.
work_yield
How much observed work still converts into structural movement.
burn_without_progress
How much activity appears after clean structural conversion weakens.
waste_score
A structural pressure score. It is not a public waste percentage.
durable_fingerprint
A recurring structural motif seen again inside the same customer workflow.
task_family_scope
Optional customer-provided labels help separate recurrence inside one workflow family from recurrence across several families.
conversion_drift
A review watch when repeated sessions show worsening structural conversion.
long_cycle_watch
A larger observer flag for rhythm across wider windows.
Install the observer. Let the memory form.
CAUM Live starts with neutral events from your existing agent workflow. The API key powers per-step receipts, session history, workflow review cards, and review-only webhooks.
Boundary
CAUM observes structure only. It does not read prompts, files, source code, messages, documents, or customer payloads. It does not judge answer truth and does not block agents.
from caum_sdk import CaumSession
caum = CaumSession(api_key="caum_live_...")
with caum.observe(
event="tool_call",
tool="repo_grep",
state_id="hash:customer_owned_state"
):
run_agent_step()One run gives a receipt. Repeated runs give memory.
Start with CAUM Receipt if you have one trace. Start CAUM Live when the same workflow repeats often enough that structural review matters during operation.