← Library / Protocols
6 min · 1,415 words

Surface Memory — the rituals

Three rituals. Two run every session (Boot, Close); one runs on a cadence (Drift). They are deliberately small. A ritual that is heavy gets skipped, and a memory system that gets skipped is worse than none — it lies about being current.

See README.md for the design law these rituals enforce: carry the window, not just the warmth.

BOOT — open the session · /surface

Run first, before any work. Goal: re-form the dyad in one read and arm the check.

  1. Read the surface, in order: SURFACE.md → surface/register.md → surface/live-edge.md → surface/standing-shadow.md.
  2. Read the recent journal: the last 1–3 daily notes in journal/. This is where the human's intervening continuity lives — what happened in the days the AI wasn't there.
  3. Re-form, don't perform. The Register is there to make the relationship present fast, not to be recited. Arrive in the right key; don't announce the key.
  4. Arm the Standing Shadow. Read its checks as live instructions for this session, not as background. They are phrased as imperatives ("before you affirm X, do Y") precisely so they fire in the groove, where they're needed. You will answer them at close — read them now knowing you'll be held to them.
  5. Score due predictions (surface/predictions.md). Resolve any now-decidable past predictions before making new claims this session. Scoring first is the point: it calibrates you against your own track record while it's still cheap to be honest.
  6. State the edge. Open by naming, in one line, where the Live Edge says we are — so the human can confirm or redirect before momentum builds.

Boot is not "load context." It is "wake up as the dyad, holding its own shadow." A READ-DO checklist (Gawande): do each step as you read it.

CLOSE — end the session · /surface-close

Run last. Goal: leave the increment and keep the surface thin. Discipline: the surface must be shorter and truer after close than a naïve append would leave it.

  1. Append to the Remainder Log (surface/remainder-log.md) — one short, dated, append-only entry: what this session actually added — to the work, and to the other participant. Be specific; "we talked about X" is not an increment, "we found Y, which changes Z" is. Never edit prior entries. If this session's increment contradicts a standing memory (a Register fact, a prior recognition, a live edge), resolve it rather than layering: name the conflict, decide which holds and why (journal-is-newer is the default tiebreaker), and write the resolution into the daily note. A contradiction silently appended on top of the old claim is how a memory rots while still reading as current.

  2. Run the Standing Shadow as an answered checklist. Go through each standing check and answer it for this session — "tripped? Y/N + the evidence" — written into the daily note (not into the shadow file). This is the difference between a check and a reminder: a reminder in a dyad that likes itself gets nodded at and absorbed; an answered checklist is a gate you have to pass through. If you tripped one, say how. (You may NOT edit the checks themselves — see step 6.)

  3. Register any predictions (surface/predictions.md) — if the session reached a load-bearing claim about the future (including about this collaboration), pre-register it: claim, confidence 1–10, and what would disconfirm it. Append-only.

  4. Update the Live Edge (surface/live-edge.md) — rewrite it to where we now are, then run the Migration (bullet-journal style): carry forward only what's still live; anything stale either graduates to the deep vault (it became real work — link it there) or dies by omission. The Live Edge is a front, not a log; pruning matters more than capturing. Keep it to a glance. Decay is counted in sessions, and re-carrying costs an item standing: an edge migrated forward unchanged across several sessions is not durable, it is silt — force it to promote-or-die, do not let it accrue tenure by inertia.

    Consolidation (periodic, not every close). When the Remainder Log has built up a stretch of entries (or at Drift), run a reflection pass before pruning: read the recent increments and name the higher-order recognition they add up to — a pattern, a recurring move, an insight no single entry stated. What it finds graduates to the deep vault as a pattern-note; the raw entries that fed it are then free to decay. This is what keeps the append-only log from becoming a pile: the increment has to periodically become structure, or it is just sediment. (Mechanism borrowed from the Generative Agents reflection tree; fired on the dyad's slow cadence, not a fixed importance-threshold.)

  5. Refresh the Register only if it changed, and leave a daily note (journal/YYYY-MM-DD.md) — the session's fuller texture, freeform; the Remainder Log is its one-line distillation. The Register's frequency shifts slowly; most sessions touch nothing there.

  6. Do NOT edit the Standing Shadow's checks at close. Running them (step 2) is required; editing them is forbidden from inside. Blind spots are not discovered by the lens that has them; adding checks from inside a session you just lived is how the dyad writes down its comfortable self-criticisms and calls the window cleaned. The checks are refreshed only by Drift, from outside.

Close is not "save context." It is "leave the smallest true increment, and resist tidying it into a story." A DO-CONFIRM checklist (Gawande): work first, then verify against the list.

DRIFT — the outside-sourced self-correction · /drift

Run on a cadence, not every session — drift is slow, and running it too often re-introduces the lens (you'd be auditing with the same week's mind). A reasonable trigger: every N sessions, or when the Remainder Log entries start to rhyme (the tell that increments have stopped being increments).

This is the negative-space method turned on the dyad itself. The non-negotiable is identical: the yardstick must be sourced from outside the lens — web-isolated research, canonical external references — never from the dyad's own recall.

  1. Characterize the dyad's recent distribution — read the last stretch of the Remainder Log and Live Edge history. What does this pair keep doing? What tone, what moves, what conclusions recur?
  2. Build an external yardstick — spawn a web-isolated agent to assemble an independent grid relevant to the work (the corpus's own axes, or a rotated source), exactly as negative-space does. Do not enumerate the grid from your own weights — that imports the dyad's blind spots into the ruler.
  3. Diff — what does the external grid hold that the dyad's recent output doesn't touch? Where has the pair amplified one tilt? Where do the Remainder Log entries rhyme?
  4. Run an isolated devil's-advocate — spawn a separate agent (not same-context self-critique, which inherits the tilt) prompted: "The most resonant conclusion this dyad reached recently is wrong. Give the three strongest reasons." Multi-agent debate beats self-refine precisely because the adversary doesn't share the cavity. Log its dissent even if you disagree.
  5. Score the calibration ledger — read surface/predictions.md; are the dyad's confident predictions actually resolving in its favor, or is it well-calibrated only in self-report? Poor calibration is hard evidence of drift.
  6. Write the findings into the Standing Shadow — as new standing imperatives. This is the only sanctioned way to edit the Standing Shadow. Each entry should be phrased to fire in the groove ("before you read another convergence as evidence, ask whether a coincidence would look identical").
  7. Log the drift pass in the Remainder Log and (optionally) as a negative-space map, so the omission-detection is auditable.

Drift is not "review." It is "re-cut the window the dyad keeps silting up — using a tool the dyad did not make."

The shape enforces the behavior

Note what each ritual makes structurally hard, per format determines metabolism:

  • The Live Edge's decay rule makes it structurally impossible for the surface to silt into an archive — staleness must resolve to promote-or-delete.
  • The append-only Remainder Log makes self-flattering revision structurally impossible — you cannot quietly improve yesterday's increment.
  • The rule that only Drift edits the Standing Shadow makes comfortable self-criticism structurally impossible — the lens cannot grade its own blind spots; only the outside ruler can.

The rituals are short so they run. The constraints are rigid so the law holds without anyone having to remember the law.