[Priority 2] Implement incremental context streaming (Context Updater pattern)

[marknutter]

Problem

Current recall uses static snapshots — load all matching entries at once. This doesn't scale well:

• Query with 50+ matches → 50+ subagent dispatches, even if top 5 answer the question
• No support for long-running iterative analyses (drill deeper based on findings)
• No dynamic refresh (swap stale entries for better ones mid-session)

Proposal (from AIGNE paper analysis)

Add ContextStream class that manages dynamic context loading:

from rlm.updater import ContextStream

stream = ContextStream(session_id='abc123')
stream.load_initial(top_entries=5)  # Start with highest-scoring

while not analysis_complete:
    findings = dispatch_subagents(stream.current_context)
    if stream.needs_refresh(findings):
        stream.swap(remove=['stale_entry'], add=['deeper_dive_entry'])
    stream.log_context_state()  # Full provenance

Three modes:

• Static snapshot — one-time load (current behavior)
• Incremental streaming — progressive loading as reasoning unfolds
• Adaptive refresh — replace stale/irrelevant fragments based on model feedback

Implementation

1. Create rlm/updater.py with ContextStream class
2. Integrate with graduated dispatch (use incremental mode automatically when >10 matches)
3. Log all context state transitions to provenance log
4. Add streaming support to recall pipeline

Impact

• Resource efficiency (only load what's needed)
• Supports long-running iterative analyses
• Better user experience (faster results for queries with clear top-ranked answers)
• Full traceability of context evolution

Effort

3-4 days

Related

• Context Updater from 'Everything is Context' paper
• Graduated dispatch (already implemented)
• Provenance logging (dependency)