feat(decay): slice 1 — surfacing-aware memory decay

[nickroci]

Summary

First slice of the LTD design from forgetting-ltd-decay-design. Deterministic, no-LLM bookkeeping that runs without curator cost.

• last_surfaced frontmatter field stamped by the daemon every time an entry shows up in priming output. Uses the per-session sent-cache as the trigger so we know when the agent has actually seen an entry. Idempotent within a day.
• 30-day decay window. last_activity = max(created, updated, last_reinforced, last_surfaced). Entries with now - last_activity > 30d AND reinforced < 2 AND not arousal_pinned: true get archived (moved to _archive/<orig-path> with status: stale + archived: <today> stamped on the moved file).
• Opportunistic sweep trigger. No dedicated thread. decay.maybe_run_sweep() is called from priming_rpc._handle_priming and from scholar.review() (both main + empty-packet paths). It self-skips unless the 24h cooldown has elapsed and no other sweep is in flight. Sweep state lives at ~/.agent-mem/sweep-state.json.
• Pure-bookkeeping mechanism, no LLM cost. The Scholar's role is curation at write-time; decay is just the operational corollary.

Failure semantics

Every public entry point is best-effort: a bad entry skips that entry but doesn't abort the sweep; a missing knowledge dir produces an empty result and still updates the cooldown; the daemon never crashes on a decay failure. Stamping last_surfaced failures get logged and swallowed in the priming hot path.

Test plan

• uv run pytest --cov — 514 passing, coverage 90.05% (up from 89.04%)
• decay module tests — 45 cases covering frontmatter mutation, sweep-state I/O roundtrip + corruption fallback, eligibility predicate corners, end-to-end archival with each guard exercised in isolation, cooldown-trigger mechanics, and the deeper defensive paths (source-unlink failure, mkdir failure, destination-resolution failure, sweep continuing after a bad entry, lock-held early-bail)
• priming_rpc bookkeeping tests — _post_render_bookkeeping called for each newly_sent entry, no-session-id branch, swallowed stamp/sweep failures
• Pre-push hook (ruff + pyright + pytest with 90% gate) green

What's not in this slice (intentionally deferred)

• encoding_strength as a write-time field set by Scholar — slice 2.
• Retrieval boost from encoding_strength — out of scope per the design's decision to keep retrieval signal coming from the cross-encoder, not from a self-reinforcing prior.
• Contradiction-reset of decay timer — flashbulb guard; slice 3.
• Resurrection from _archive/ on new contradiction — slice 3.
• Drift-detection reconsolidation — slice 4.
• Retrieval-induced forgetting (RIF) for near-duplicates — slice 1b, to land after this one so we can see decay behaviour first before stacking the suppression mechanism on top.

🤖 Generated with Claude Code