Store a financial fact, observation, or decision in persistent memory. Always provide event_time_iso as when the event occurred, not now. Add ticker/metric/entity metadata for precise supersession detection — this lets Lians automatically replace stale guidance numbers.

Retrieve the most relevant CURRENT memories for a query. Returns only presently-valid facts — superseded facts are excluded at the DB layer. Call this before answering any question that may be in memory. Use filters={ticker: NVDA} to narrow to a specific instrument.

Retrieve memories that were valid at a specific past point in time. Use for compliance and audit: 'What guidance did we have on 2026-03-01?' Later superseding updates are excluded — this is true point-in-time recall. mem0 and Zep have no bitemporal model with compliance audit stack.

Reconstruct the complete memory state and full audit event trail at a past point in time. Returns every memory valid at as_of plus the timestamped, hashed event log behind them. Use for regulatory audit submissions and trade reconstruction.

List open conflict flags — cases where two sources reported different values for the same fact at the same event_time. Use this to surface data quality issues before they affect decisions. Returns up to 20 open conflicts with both memory contents so a human or LLM can decide which source to trust.

Return the full supersession history of a memory — every prior version of the same fact and the chain of updates that led to the current value. Use when asked 'how did this guidance number evolve over time?' or when investigating why a memory was replaced.

Return every recorded version of a structured fact ordered by event_time. Query by ticker + metric — ideal for time-series views like 'show me how AAPL EPS evolved over the last four quarters'. Superseded versions are included so you can see the full revision history. Entity normalization: 'Apple Inc.', ISIN 'US0378331005', and 'AAPL' all resolve to the same fact series automatically.

Detect lookahead bias in a backtest simulation. Scans the agent's memory store and flags every fact the agent couldn't have known at the given simulation date. Returns FUTURE_EVENT (event_time is after the checkpoint) and LATE_REVISION (the revised figure hadn't been published yet) flags. A clean report (is_clean=true) is the proof a risk committee needs.

Record whether a recalled memory was helpful, incorrect, outdated, duplicate, or ignored. Negative feedback flags evidence for review and never silently deletes it.