The memory writer layer is the first real implementation of ClaimFlow AI workflow memory.
Its job is not to retrieve memory yet. Its job is to write safe, reusable memory cards from past workflow outcomes.
Flow:
past workflow outcome
→ normalized memory observation
→ WorkflowMemory card
→ MemoryUpdate audit row
By the end of Day 3, ClaimFlow can create long-term memory records from:
memory-observations.json
workflow-memories.seed.json
real ReviewDecision rows
correctedJson diffs
memory-observations.json ismemory-observations.json is the bridge between old workflow history and real memory.
It stores normalized observations from past ClaimFlow outcomes, such as:
human correction
prior rejection
prior review decision
agent action history
claimant pattern
vendor pattern
policy-history signal
recurring workflow pattern
Example:
Reviewer corrected missing policyNumber for CUST-W5-001.
This becomes an observation with:
observationId
sourceType
sourceId
entityType
entityId
fieldPath
beforeValue
afterValue
riskLevel
recommendedMemoryKind
summary
safeUse
mustNotDo
evidenceJson
The important idea:
old claim data is not dumped directly into agent context
old workflow history is first converted into compact, safe memory observations
workflow-memories.seed.json is