David F GliddenandClaude Opus 4.8
3f9a89b00c
Basic Memory v0.21.6 first sync over the live memory dir (steward-authorized live-dir trial, Option A 2026-06-06): adds permalink: to frontmatter, refolds long YAML description lines, strips final newlines. Bodies untouched — verified via full diff classification. From this commit forward, any diff in claude/memory shows only what Basic Memory or the session writes. Trial design: MemPalace untouched as incumbent; git status check on this dir at every wrap; end-of-day evaluation (recall quality, sync robustness, rebuild-from-files, malformed-file behavior). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
6.3 KiB
name, description, type, originSessionId, permalink
| name | description | type | originSessionId | permalink |
|---|---|---|---|---|
| Onboarding vs steady-state lag — threshold-governance frame | Slowness in L1 ingestion is not a bug-or-not-bug binary; it is a threshold-governance question. Two regimes (onboarding / steady-state lag) collapse in current code but are different in kind. Defining the acceptable envelope is L2-shaped work, not engineering verdict. | feedback | a3ddbf34-da48-4129-acba-658f2f869783 | claude-memory/feedback-onboarding-vs-steady-state-thresholds |
L1 slowness as threshold-governance, not bug
When evaluating L1 ingestion or recall slowness, do not frame it as "is this fast enough or not." Frame it as: which regime is this, what is the acceptable envelope for that regime, and is the observed behavior inside or outside the envelope?
Why: Steward articulated this as a load-bearing reframe during the 2026-04-19 deep audit, in response to my drift toward extrapolating a drain-rate ("19 days at this rate") as if it were a verdict. He explicitly named that slowness up to a defined point is acceptable — what's missing is the definition of the point. Without the envelope, both panic and complacency are equally ungrounded.
Verbatim quotes (steward, 2026-04-19):
"If this is to be a slow process (either in onboarding or recovery), that can be accepted up to a certain point. That point needs to be defined of course."
"Kai is 9 months old tomorrow (9 months in the belly, 9 months on the belly...). when I consider the time it takes a human to learn the necessary basics for survival, social integration, basic foundation knowledge — onboarding — it is a long and time-expensive process. A user of CapableMind needs to know and accept (up to the yet undefined point) that ingesting one's information to be able to use it without thinking or having to teach your tools everything will be necessary time-expensive."
"Idem for recovery: a human who has a catastrophic accident which requires 'finding oneself', or recovering their memory (if possible), or, worst case scenario, needs to relearn must have unlimited patience. Same for CapableMind replay. But again, up to that yet undefined point."
The two regimes (different in kind)
| Regime | Character | Slowness reading |
|---|---|---|
| Onboarding | One-time, large backlog, ground-acquisition | Legitimate time-cost; operator should know and accept |
| Steady-state lag | Live events falling behind; cursor drift under load | Pathology; threshold breach |
L1 currently collapses these — modules in both states report ready and system_status: healthy. GH issue #139 ("Distinguish 'catching-up-from-zero' vs 'catching-up-from-lag' in system_status") gestures at the engineering fix but does not yet solve the constitutional shape.
How to apply
- Forensic work: produce data that lets the steward + jurist define the envelope. Not endpoints — distributions. Rate-of-drain over time, cost-per-entity vs graph density, attention-cost per ingestion unit, recall-quality curve over time.
- Diagnostic claims: do not assert "too slow" or "fast enough" in written findings. Surface the regime, the observed behavior, and the question of where the envelope sits. Let the threshold-decision be the steward's.
- Recovery work: same pattern in reverse — unlimited patience within bounds that remain to be named. Includes: how much of the prior substrate survives catastrophe, at what rebuild cost.
- Operator experience: if the steward has to intervene N times during onboarding, the threshold is already crossed — even if eventual convergence is fine. Attentional cost is part of the envelope.
- Recall quality curve: does the system become useful incrementally (infant learning) or only at completion (useless until done)? This is itself an envelope marker.
Candidate marker dimensions (surfaced as inputs, not decisions)
- Temporal — acceptable onboarding duration in absolute time
- Attentional — operator interventions per ingestion unit
- Epistemic — recall quality available at T+N for varying N
- Recovery-specific — survivability + rebuild cost after catastrophic event
Important calibration — the frame is not a license
The reframe defines what acceptable means; it does NOT license accepting current state. Steward calibration (2026-04-19, after the initial reframe):
"It is one thing for Seb and I to accept a slow onboarding or recovery — we're the creators and have patience that end-users won't. But I thought the reflection was important to share with you. The two processes need to be as fast as is possible of course. The projected completion time for my machine is untenable (not to mention probably not accurate either) so this process must be sure improved."
So: the threshold-governance frame is the analytic shape for the work. The working assumption is that current state is outside the envelope, not inside it. Forensic findings should produce both:
- The shape of data needed to define the envelope (distributions, regime markers, recall curve).
- The shape of data needed to improve — bottleneck localization, scaling-law identification, architectural alternatives.
End-user patience is the binding constraint, not creator patience. Any improvement target should be calibrated against what an end-user would tolerate, not what the steward and Seb will tolerate as part of the build.
Anti-pattern to avoid
The drift caught here was: stating a linear-rate extrapolation ("83 events / 8h → 19 days to drain backlog") as if it were a verdict on system health. Even when hedged ("at this rate"), that framing imports the bug-or-not-bug binary. The threshold-governance frame demands a different output shape: distribution, regime identification, and surfacing the envelope question — not a single number presented as judgment.
Connection to L2
The two-regime distinction is constitutional in nature, not engineering. It connects to:
- DN-GOV-05 (bounded self-repair) — the four conditions are envelope conditions
- DN-GOV-06 (temporal authorization shift) — threshold crossing as authorization trigger
- ICP-19 / external review — envelope-defining requires external grounding (earned confidence)
- The horizon work on
ground-up-model-horizon— substrate trust is itself a function of patience-under-defined-bounds