v2.15.0 — Memories can finally mature

Correctness release, and the most consequential so far for anyone with a brain more than a few weeks old. Two defects meant the system reported health it did not have while quietly failing to do its central job.

⚠️ Your purity score will drop ~23 points. That is the fix working.

metric	before	after	after optional alias cleanup
connectivity	1.0 (saturated)	0.085	0.085
diversity	0.256	0.256	0.948
purity	—	−23	+14

Connectivity counted alias edges — dedup pointers with weight 0.0 carrying no meaning. On a real brain they were 89.4% of all synapses (137,871 of 154,252), so the metric measured plumbing, not knowledge. Stored purity/grade from before this release is not comparable with after — treat the upgrade as a new baseline.

Memories never matured

Fiber.create mints a dash-form uuid4; storage folds ids to underscore form; the maturation table stored whichever the caller passed. The rows never joined.

Of 1,920 live maturation rows, 1,277 were dash-form — and not one had ever been rehearsed or promoted. Every rehearsed row (77) and every semantic row (9) was underscore-form. get_maturation returned None, rehearse() never ran, the EPISODIC→SEMANTIC spacing gate could never be met, and pattern extraction was blind to the same 1,277 fibers. A blanket except Exception had logged the resulting collisions as "Skipping maturation save" for months.

Lookups now resolve through the record id — canonical in both eras — so legacy rows heal on their next write, no migration required. Expect the first consolidation after upgrading to do noticeably more work.

Decay compounded on itself

Nothing recorded that a synapse had been decayed, so every run re-measured from created_at and re-multiplied an already-decayed weight — quadratic in run count. Ten daily runs applied ~55 days of decay, not 10.

Decay now bills only the interval since _last_decayed. The exponents telescope, so decay per day of wall-clock time matches the documented Ebbinghaus curve. Weights fade slower than in previous builds — the compounding was the bug, not the cure. Rows with nothing to bill are skipped, removing ~57k pointless rewrites per run.

Also fixed

• Dedup re-inserted the same alias edge every consolidation — 2,375 pairs had produced 144,565 rows, growing ~40k/day. Deterministic per-pair ids make it idempotent.
• Connectivity contradicted itself across smem health, smem_stats, the maintenance handler and topology analysis. One shared definition now.
• Knowledge surface reported "not generated" depending on your shell's directory — $HOME was classified as a project root because it holds the global .surrealmemory/.
• Warning texts collided units — "connectivity 1.0 (target 3-8)" pairs a 0-1 score with a raw ratio, reading as "far below" when the raw value was 13.5, far above.
• LOW_DIVERSITY no longer fires on brains using many synapse types. The live brain uses 17; the warning claiming "1 of 8" was reading an empty scope.

Full changelog: v2.14.0...v2.15.0