feat(orchestrator): memory architecture — the axes §7 does not have

[sshlg]

SKILL.md §7 gives four memory layers and their lifetimes. Lifetime is one property of a memory system, not its taxonomy — and the survey this file pins says so directly: long/short-term is "proven insufficient to capture the diversity and dynamics of contemporary agent memory systems." Short- and long-term behaviour emerges from how often formation, evolution and retrieval run, not from which box a thing was filed in.

Three gaps, each measured against what the pack already shipped

Retrieval had 2 mentions across 1,938 lines of references. It is four decisions, not one: whether to retrieve at all and from which store · what query to retrieve with · which strategy runs the search · what actually reaches the prompt.

The first matters most and is the one nobody instruments. An agent that overestimates its own knowledge and skips retrieval produces a confident answer built from nothing — no error, no empty result, no latency spike. Invisible to every check that watches for failures.

Facts were not separated from experience, and facts split again by entity. Layers 3 and 4 are experiential — earned from contrast (§8). Nothing in them is a factual store, and a stale fact about the user makes the agent rude while a stale fact about the environment makes it wrong. One expiry rule for both is wrong twice.

Forgetting was decay only. Frequency-based eviction is the easy one to reach for and the one that eliminates long-tail knowledge — the rarely-read entry that prevents the rare expensive mistake. Where storage is not the binding constraint: demote, not delete.

Also carried

• Forms other than token-level, with their costs: parametric memory cannot be selectively deleted (collides with a right to be forgotten); latent memory cannot be inspected.
• Shared memory ≠ agent-sync. Leases decide who may write a file now; they say nothing about whose experiential memory is trustworthy enough to act on.
• Trustworthy memory, three pillars. Abstention under low-confidence retrieval needs a similarity floor to be expressible at all — semantic search always returns K results, so an empty store and an irrelevant one look identical to the model.

What it does not claim

No benchmark numbers reproduced. The frontier sections are frontiers, not practice. It was written from the survey, not from running these systems — where the pack has its own measured experience it stays marked as the pack's.

Discoverability, measured

Against the query "agent memory architecture: forms, functions, dynamics, retrieval, forgetting" this skill ranked 8th among installed skills and now ranks 1st — four distinct term hits against two for the next entry. Description moved 964 → 965 chars, inside the 970-char working limit this repo enforces below the 1024 cap.

(The first attempt to measure this reported the rank getting worse. That was a broken extraction regex returning an empty description, not a real result.)

Source pinned with a read date: arXiv:2512.13564v2, 13 Jan 2026 · read 2026-08-27.

🤖 Generated with Claude Code