ADR-0050: Temporal / Confirmation-Weighted Recall — Un-Collapsing Resonance and Recency

Status: Accepted — mechanism measured and supported (L8, 2026-07-26), shipping OFF by default. The half that handles contradiction is blocked on a writer, not on ranking (see Consequences). Relates to: ADR-0035 Cap 8 (the effective_at / observed_at / expires_at temporal-truth fields this ADR is the first ranking consumer of), ADR-0048 (KANNAKA_RECALL_ENERGY_EXP — same shape, same off-by-default discipline), ADR-0049 (facet encoding — reach; this ADR is ordering, and only ever re-orders a pool that encoding made reachable). Evidence: experiments/results-L8.tsv, experiments/notes/2026-07-26-L8-temporal-confirmation-recall.md.

Context

The question came from outside the ladder — a reply to the Product Hunt announcement on Mastodon (@aiappsapi, 2026-07-26):

"does recall strength depend on how recently a memory was confirmed, or only on how strongly it resonates? Those two come apart quickly once stored facts start contradicting each other."

The answer was resonance only, and the code was unambiguous about it:

• Hemisphere::resonate_with_energy_exp ranks similarity × energy^e and reads no timestamp at all.
• energy is pumped by access, not confirmation — the rich-get-richer bias ADR-0048 measured.
• The ADR-0035 temporal fields are persisted but read only as a post-recall filter in swarm brief (bin/kannaka.rs:3622), never in ranking.
• The ~693-day amplitude decay (medium/core.rs:54) is on the FLAT path; the default CHIRAL path is age-blind.

So two axes that come apart under contradiction were collapsed into one. A superseded fact and the fact that replaced it are near-identical text — they land side by side in the candidate pool, and similarity separates them arbitrarily.

Decision

Add a third factor to the ranking product, off by default:

resonance = similarity × energy^energy_exp × tweight^temporal_exp

hemisphere::temporal_weight(meta, now, half_life_days, floor) folds two things:

1. Truth status — past expires_at (superseded) or before effective_at collapses to a floor, never to zero.
2. Confirmation recency — a currently-true fact decays 0.5^(age/half_life) from observed_at, falling back to created_at, clamped at the same floor.

Flags: KANNAKA_RECALL_TEMPORAL_EXP (default 0.0 = off, byte-identical), KANNAKA_RECALL_TEMPORAL_HALFLIFE_DAYS (180), KANNAKA_RECALL_TEMPORAL_FLOOR (0.25).

resonate_with_energy_exp now delegates to a new resonate_with_weights with temporal_exp = 0.0, so every existing caller and test is unchanged by construction.

Applied at BOTH scoring points — the full-corpus pass and the xi re-rank — so it governs which candidates enter the 2k pool. This is the ADR-0048 review's lesson carried forward: post-fetch re-ranking cannot promote what was never fetched.

The floor is never zero. Discounting a superseded fact is the point; making it unretrievable is amnesia with a flag. An agent has to be able to answer "what did we use before".

Measurement (L8)

New research level --level 8. The fixture is built so the two axes must come apart: fact families are near-identical sentences differing in one value token, all answering the same cue, where version v expires_at exactly when version v+1 was observed_at. Similarity cannot separate them; only time can. 10 corpus seeds, each arm measured paired against its own baseline on the same corpus.

Gates were pre-registered: P1 current-fact retrieval improves (> 0.5), P2 uncontradicted facts not degraded (≥ 0.95), P3 superseded facts stay reachable (≥ 0.99), P4 instrument-live + default-byte-identical (hard PASS/FAIL, not folded into fitness — a dead mechanism must fail loudly rather than score well).

At exp = 0.25, floor = 0.25, half_life = 180d: current-fact MRR 0.4774 → 0.8503 (Δ +0.3729, 71% of available headroom), P2 = 0.9667, P3 = 1.0000, fitness 0.151 vs 0.500 baseline. Revert-and-confirm-fail holds — at exp = 0.0, P1 is exactly 0.0000.

Two findings outrank the win itself:

The floor is the amnesia knob. Floor 0.05 evicts the past (P3 0.9399, fails the gate); the safe plateau is 0.15–0.50 (P3 = 1.0000).

Freshness, not supersession, is most of the win. Across a 120× half-life range (30d → 3650d) P1 moves only 0.699 → 0.743. At 3650d the recency term is effectively inert and the supersession floor alone still delivers P1 = 0.7425 — but P3 collapses to 0.9356. Both sub-mechanisms independently produce the win and differ entirely in what they cost: freshness-promotion is nearly free, supersession-demotion is what risks the past. Keep the floor generous; let recency do the work.

Consequences

The blocker moved from ranking to writing. Code-verified: outside tests and the L8 harness, the only writer of the three temporal fields is HrmStore::set_temporal, reachable from exactly one place — kannaka remember --effective/--observed/--expires (bin/kannaka.rs:1557). Nothing auto-populates them. On today's live HRM:

• observed_at is None everywhere → recency falls back to created_at and still works, as age-since-storage.
• expires_at is None everywhere → the supersession half is completely inert.

Enabling the flag on the live corpus today therefore buys age-decay ranking, not contradiction handling. That is the honest scope of what this ADR ships.

Follow-up, and it is the harder half: something has to write supersession down. Deciding that a new fact replaces an old one is a judgement, not a timestamp — it needs its own design (and its own adversarial review) covering what counts as a contradiction, what happens on partial disagreement, and who is allowed to expire a memory. Single-writer discipline applies: only the writer may stamp; readers rank.

Not addressed here: auto-bumping observed_at on re-assertion (the "confirmation" the original question named), inference of supersession, and any interaction with dream consolidation, which may merge versions before ranking ever sees them.

Risk accepted: the mechanism is validated on a synthetic fixture with clean, single-attribute contradictions and disjoint vocabulary. Real contradictions are messier and partial. This ships off by default, and the L8 harness stays in the ladder so any future change to the ranking path is re-measured against the same pre-registered gates.