Why

BEAM Sprint 2 dry-run (iter 7 v3) is leaving temporal-reasoning (TR) and multi-session-reasoning (MR/MSR) probes on the table: TR 1/2, MSR 0/2 in the latest run, well below the Honcho TR=0.644 / MR=0.631 reference. Raw created_at timestamps don't match query phrasing like "recently" or "last week"; the retrieval path needs a coarser, scale-invariant signal it can match against.

Log-spaced recency bins are the cheapest first cut at this — Blueprint M5 calls out "AtomicMemory has naive timestamps … No Laplace bank, no drift vector, no scale-invariant representation"; log bins approximate Laplace at ~5% of the cost.

What

EXP-12 from phase2-implementation-plans-2026-04-29.md. Two pieces:

• Ingest-time tag: every fact written by storeProjection carries metadata.recency_bin (a debug breadcrumb; retrieval recomputes from created_at).
• Retrieval-time boost: new applyRecencyBinBoost stage in the search pipeline runs after applyCurrentStateRanking and before applyConcisenessPenalty. When the query has a recognizable recency marker, the stage adds recencyBinBoostWeight × computeBinAffinity(queryBin, factBin) to each candidate's score (1.0 for exact bin match, 0.5 for an adjacent bin, 0 otherwise) and re-sorts. Skipped entirely when current-state-ranking already triggered, to avoid double-counting two recency-flavored signals.

Bin schema (and why log-spaced)

1m    →   ≤ 60 s
10m   →   ≤ 10 min
1h    →   ≤ 1 h
10h   →   ≤ 10 h
1d    →   ≤ 24 h
10d   →   ≤ 10 days
100d  →   ≤ 100 days
older →   anything older

Constant ratio (~10×) per rung. Adjacent bins are close in human terms ("yesterday" vs "last week" feel related), so we award half-credit for adjacency. BEAM conversations span minutes-to-months simulated; a fixed-stride scheme would either lose minutes-scale resolution or balloon the bin count.

Config-override copy-paste (defaults-off)

{
  "config_override": {
    "recencyBinBoostEnabled": true,
    "recencyBinBoostWeight": 0.10
  }
}

Both keys are in INTERNAL_POLICY_CONFIG_FIELDS. Env-var equivalents are RECENCY_BIN_BOOST_ENABLED and RECENCY_BIN_BOOST_WEIGHT.

How it composes with EXP-05

Independent. EXP-05 touches extraction-enrichment.ts, memory-service-types.ts, memory-storage.ts:96 (instruction-tag), and search-pipeline.ts:755-761 (instruction boost). EXP-12 touches temporal-fingerprint.ts, temporal-query-expansion.ts, recency-bin-ranking.ts (new), memory-storage.ts:96 (recency-bin breadcrumb), and the same protection-stage block in search-pipeline.ts. Both share that protection-stage block but write distinct stages — instruction-boost goes between current-state-ranking and conciseness-penalty; recency-bin-boost goes immediately after current-state-ranking. The RESERVED_METADATA_KEYS set is the only collision point — EXP-05 added fact_role, EXP-12 adds recency_bin. Both keys are pre-allowlisted for the metadata drift guard.

Test plan

• src/services/__tests__/temporal-fingerprint.test.ts — bin assignment table tests (boundary cases at 1m, 10m, 1h, 10h, 1d, 10d, 100d, plus future-dated clamp) and computeBinAffinity (exact / adjacent / non-adjacent).
• src/services/__tests__/recency-bin-ranking.test.ts — boost applied when enabled, no-op on weight=0, no-op on unrecognizable queries, short-circuit when current-state-ranking already triggered, weight applied as configured, recompute-from-created_at (ignores stale persisted hints).
• src/__tests__/reserved-metadata-keys.test.ts (drift guard) — passes with 'recency_bin' added.
• src/__tests__/config-partition.test.ts — passes with new keys in INTERNAL_POLICY_CONFIG_FIELDS.
• npx tsc --noEmit — clean.
• BEAM TR / MR / MSR slice with recencyBinBoostEnabled: true — pending dry-run iter on feature/exp-12-log-spaced-recency-bins.

Rollback

recencyBinBoostEnabled: false (default). The persisted metadata.recency_bin field is harmless metadata; no migration is needed if the flag is disabled or the feature is reverted.