awesome-loop-engineering / examples /agent-memory-lifecycle-loop.json
cy0307's picture
Grow implementation kit: 22 patterns, 22 contracts, 671 resources
bbb3e00 verified
Raw
History Blame Contribute Delete
3.89 kB
{
"name": "Agent memory lifecycle loop",
"objective": "Carry governed operational memory between recurring agent runs: explicitly written, scoped on recall, audited, consolidated, and deletable.",
"trigger": {
"type": "scheduled",
"cadence_or_event": "Weekly consolidation pass, plus a session-end candidate-write hook after each run."
},
"intake": {
"sources": [
"candidate records from recent runs",
"live memory store with provenance",
"retention and privacy policy",
"audit log",
"recall-scope definitions"
],
"selection_rule": "Process only candidate records written since the last pass and live records due for audit, expiry, or review."
},
"workspace": {
"isolation": "Staging area for candidate records; the live store changes only through gated promotion.",
"allowed_actions": [
"distill run artifacts into candidate records",
"deduplicate and merge with provenance",
"audit for staleness, contradiction, and scope creep",
"apply logged expiries and redactions",
"run recall-scope test queries"
],
"disallowed_actions": [
"store raw transcripts or credentials",
"delete audited records silently",
"write to the live store without promotion",
"broaden a record's recall scope without review"
]
},
"context": {
"required_files": [
"memory store manifest",
"retention and privacy policy",
"recall-scope map",
"audit log"
],
"runtime_sources": [
"session-end run artifacts",
"quarantine queue",
"recall test results"
]
},
"agents": [
{
"role": "recorder",
"responsibility": "Distill each finished run into explicit candidate records instead of raw transcripts."
},
{
"role": "auditor",
"responsibility": "Flag stale, contradictory, sensitive, or unattributed records and quarantine failures."
},
{
"role": "consolidator",
"responsibility": "Merge duplicates, promote repeated lessons to principle-level records, and retire superseded ones."
},
{
"role": "recall gate",
"responsibility": "Scope what each future run may read by task and access class, and verify scopes with test recalls."
}
],
"verification": {
"gates": [
"every record carries provenance, scope, and an expiry or review date",
"no raw transcript, credential, or policy-barred content in the live store",
"contradictory records cannot coexist unresolved",
"recall tests retrieve intended records and nothing out of scope",
"deletions and redactions are logged and reversible for the review window"
],
"receipts": [
"consolidation receipt linking merged, promoted, and retired records",
"audit log entries for every write, merge, prune, and redaction"
]
},
"state": {
"artifacts": [
"memory store with typed records",
"audit log of writes, merges, prunes, and redactions",
"consolidation receipts"
],
"update_rule": "Only promoted records enter the live store; quarantined and expired records never recall; every mutation appends to the audit log."
},
"budget": {
"max_retries": 1,
"max_runtime_minutes": 60
},
"escalation": {
"conditions": [
"sensitive or ambiguous records",
"contradiction between operator instruction and remembered preference",
"retention-policy questions",
"records whose provenance does not trace to a legitimate run"
],
"destination": "Review queue with the quarantined records and their provenance evidence."
},
"exit": {
"success": "All candidates processed, audits clean, recall-scope tests pass, and the receipt is written.",
"stop_without_success": "Churn cap reached or policy failures found; the live store keeps its last audited state."
}
}