[Priority 1] Add provenance logging and lineage tracking

[marknutter]

Problem

We can't answer:

• 'How was this memory created?' (trace back to original session)
• 'What recall sessions used this entry?' (forward lineage)
• 'Replay this recall session with exact context state'

No systematic logging of transformations (archive, summarize, recall, validate).

Proposal (from AIGNE paper analysis)

Add provenance logging for every transformation:

• archive — session → memory entries
• summarize — full transcript → summary entry
• recall — query → retrieved entries → subagent findings → answer
• validate — answer → confidence score → human override

Schema

CREATE TABLE provenance_log (
    event_id TEXT PRIMARY KEY,
    timestamp REAL,
    event_type TEXT,  -- archive, summarize, recall, validate
    source_ids TEXT,  -- JSON array of input entry IDs
    output_ids TEXT,  -- JSON array of resulting entry IDs
    agent TEXT,       -- Which component performed this
    metadata TEXT     -- JSON with confidence, reasoning, etc.
);

CLI

• rlm provenance <entry_id> — show full lineage
• rlm provenance --replay <session_id> — reconstruct exact context

Impact

• Debugging (why did this recall fail?)
• Reproducibility (exact replay)
• Accountability (audit trail)
• Trust (transparent transformations)

Effort

1-2 days

Related

• Provenance/traceability from 'Everything is Context' paper
• Context Evaluator (needs provenance for validation)