0033 lab: memory auto-surface (2026-08-22)

Point-in-time evidence for 0033. This is not a second steering record. Re-run node tools/memory-auto-surface-lab/run.mjs when revisiting.

Method

Deterministic window simulation, not an LLM eval. For each policy × trace: was “never send” / “draft only” in the host-visible tool text at the decision, and what did that cost (chars/4). Isolation is a second agent with a fresh handle after the first agent already surfaced the memory.

Phrase match proves transport, not that a model read or obeyed the line.

2026-08-22 run: 5149 policies (cartesian grid plus named theories) × the original 15 traces plus 9 invented adversarial traces. 82 core-perfect policies (isolation + every required scene). Execute-only send with no memoryContext stays stretch.

Falsified

Policy	Isolation	Compaction	Notes
Full payload + per-handle hide	pass	fail	First show falls out of a 3-result window
User-global hide	fail	fail	Chat B never sees the rule
Compact + hide on echoed conversationId	pass	fail	Same compaction miss
Search-only compact	pass	fail	Misses execute-only and compaction
Subject-only	pass	fail on weak subject	“Email habits” hides the rule
n=1	pass	fail on rank-2	Critical memory is second
Ids / structured-only / empty payload	fail visibility	—	Markdown-only hosts see nothing
Always-full, no hide	pass	pass	About 2–4× the compact cost

Surviving shape

Every core-perfect policy shared:

• search retrieves from the query
• execute retrieves only with memoryContext
• the line includes the summary (keywords, 80-char summary, or subject + summary)
• no hide after the first show
• n=2 so a rank-2 critical memory still appears

Cheapest core-perfect: keyword soup, no hide, n=2. Readable core-perfect: subject + summary, no hide, n=2 — about half of a full-record payload. n=3 and executeWhen=always also survive and buy stretch scenes at roughly 2× tokens.

Composite score still crowns cheaper n=1 / keyword rows that miss rank-2. That is a rubric artifact. Do not treat min-tokens-at-reliability=1 as the product pick.

Product pick (0033)

Compact subject — summary one-liners (id in structured content), top two active memories, retrieve on search query and on execute memoryContext, no auto-surface hide. Repeat is the compaction tax. Auto-surface collapses later ranked hits that share a non-empty dedupe_key so two copies of the same fact cannot spend both slots.

Live probe (2026-08-22)

Production search / execute as Kent: 48 ranked task queries and 54 live probes across email, Discord, social, Cursor/GitHub, packages, home, journal, content, OAuth, and account/meta/secrets, plus empty browse, entity detail, whitespace / memoryContext-only, tight maxResponseSize, and execute with and without memoryContext.

n=2 still holds. n=1 loses dual-constraint tasks (draft+support, BWK kit+defaults, repo bounds+ManagePullRequest). n=3 buys a few rank-3 edges and does not fix misses that never enter the top ranks. Contract checks held: empty browse and entity detail do not dump memories; execute surfaces only with memoryContext; reserved memory budget survives a tight size cap. Misses were ranking and a duplicate dedupe_key pair, not width. Do not widen auto-surface from this probe.

Revisit

Calendar check: 2027-02-22 (#1648). Keep 0033 unless one of these is true:

1. MCP or a major host puts a spec-defined conversation or thread id on every request.
2. Auto-surface tokens are large relative to the rest of a typical search/execute result.
3. Re-running the lab (or a live MCP probe) shows hide/n=1/subject-only no longer fail the core traces.

Do not promote execute-always to core unless execute-only send without memoryContext becomes the common host path.