TIER 0P1 · now
Working memory
- scope
- within one run
- what it forgets
- what it already sent this run
- the bill
- cost is O(n²) in task length
The field evicts old pages, then re-renders the current one in full. Rote sends the diff — built and measured.
Every harness has memory.
None of them manages it.
the memory manager for browser agents
Browser agents treat the context window as a garbage dump — append, and hope. Rote treats it as a managed resource: a budget, an eviction policy, a layout contract, and a trust gate on the way back in.
- 37.2%
- slower token growth
- 75/75
- verified successes
- 99.6%
- smaller observations
An ordinary agent
re-reads everything, every step
pages re-read so far0
The same task with Rote
keeps notes, reads only what changed
pages read so far0
docs/01 · the problem
The window nobody manages
The naive loop re-sends its whole history every step — 1 + 2 + … + n prompt-units, O(n²) in task length. The field's fix is to evict old observations, and the major harnesses now do. But every one of them still re-renders the current page in full, every step. No major harness sends the diff.
fixture B2 · input tokens per call · frozen pages
step 1
637
step 2
677
step 3
716
step 4
759
step 5
800
step 6
839
step 7
876
step 8
917
step 9
953
+38% per-call growth over ten steps — and 21% of the run's input bill is re-reading text it already sent, on a page that distills to ten nodes.
The context window is RAM. Observations are pages, dropping them is eviction, diffing is delta encoding, the prompt cache is L2, compaction is GC. Every one of those has a manager in an OS. None of them has one in a browser agent.
Rote's tier-0 policy is one sentence: keep what you did, not what you saw. The action ledger survives — ~37 tokens per step. Stale observations leave the window, and the current page arrives as a diff against the last grounded bootstrap.
The name is the thesis: rote — doing something from memory, by repetition, without re-deriving it. Memoization applied to agent trajectories, invalidated by assertion rather than TTL.
docs/02 · the memory spine
Three amnesias, three tiers
Browser agents forget at three timescales and pay again at every one. Each tier is a memory Rote manages — and every tier is assertion-gated on the way back in, because memory that might be wrong is worse than no memory.
TIER 1designed
Episodic memory
- scope
- across runs of a task
- what it forgets
- the procedure that worked yesterday
- the bill
- run #50 costs what run #1 cost
The field ships unverified replay. Rote adds the trust gate. (P2)
TIER 2designed
Semantic memory
- scope
- across tasks on a site
- what it forgets
- how the site behaves at all
- the bill
- every task re-learns the portal
Nobody. Site memory is designed, not built. (P2)
trust gateNot a fourth tier — the precondition for all three. Reuse without verification is a machine for repeating a mistake at volume.
docs/05 · tier 0
Four levers on the curve
The field ships eviction alone; Rote manages the whole window. Three of the four levers are live; the honest ledger says so about the fourth.
A11Built
Observation eviction
kills the dominant quadratic term
Prior observations leave the window; the action ledger stays. Growth drops from ~172 to ~37 tokens per step.
A4Built · measured
Diff observations
~90% off the constant
After a grounded bootstrap, each step sends a diff. In G1: 849 diffs, median 24 chars, median render-size reduction 99.6%.
B3Built · qualified on OpenAI
Cache-layout discipline
discounted billing on the surviving prefix
Nothing above the stable line may ever mutate. The immutable prefix routes through prompt_cache_key — a 20.5% cost cut at WP-N25.
B4Not built
History compaction
turns the curve from quadratic to linear
Summarize the far tail of the action ledger under budget. The one tier-0 lever still on the bench — deliberately deferred to P2.
T10 · exit gate G1 · PASS
The curve, measured
Cumulative logical input — uncached input + cache reads + cache writes, so provider caching cannot masquerade as memory reduction. gpt-4.1-mini, 15 matched repetitions per cell, 75/75 verified successes per harness, 95% seeded-bootstrap CIs.
37.2% slower growth [95% CI 35.6–38.8] vs a 30% launch floor
Cumulative logical input per task, by task length
Browser Use 0.13.6 Rote
the honest ledger
In this frozen matrix Rote was not cheaper: 5.4% more expensive at WP-N25 with 6.4% higher p50 latency, because Browser Use received more discounted cache reads. That gap is what the run below was made to close — and on the shortest cell, WP-N09, Rote still loses on billed cost. A long-task win, not a universal one. Every claim carries its receipt in the run reports.16.0% cheaper than Browser Use at WP-N25
T11 · mean billed cost per task, before → after cache-key routing
Rote, before cache key Rote, after Browser Use
the sacred invariant suite
Five invariants, encoded in tests
Non-negotiable, never “just this once.” Every one is enforced by the sacred invariant suite that touches every executor exit path.
| № | Invariant | The rule | Enforced by |
|---|---|---|---|
| I | Never silently wrong | No code path may report success when a verify or expect check failed — success is decided by page state, never by the absence of an exception. | The sacred invariant suite touches every executor exit path. |
| II | Never worse than baseline | Fallback to the plain agent is always reachable and clean; a Rote miss costs one cheap match call. | Fallback paths log why they fired — classification, not just that. |
| III | Never cross environments | A playbook learned on staging can never fire on prod. | The structural fingerprint is a hard gate before any fuzzy matching. |
| IV | Everything versioned | Store mutations are append-only — no in-place edits of playbooks or patches, ever. | Auditable, diffable, exportable as human-readable YAML. |
| V | Every model call tagged | All usage flows through one client wrapper with a source tag — planner, matcher, slot, repair, verify, distill. | Direct SDK calls outside the wrapper fail lint. |
docs/05 · roadmap
From a managed window to a substrate
No number, no launch — and the next number is G2's.
One phase at a time.
Each behind a gate with a number on it.
P0done
Foundations
Recorder, verified replay, the sacred invariant suite.
shipped
P1← we are here
Working memory
“The first browser agent with a managed context window.”
2026 Q3
G1 · the curve✓ 37.2%
G2 · the levelpending
P2planned
The harness that learns
“Your 50th task on a site costs a fraction of your 1st.”
2026 Q4
P3planned
Speculation
“Warm flows bounded by think-time only.”
2027 Q1
P4planned
Fleet & enterprise
“10K tasks a day, audited, lowest $ per task.”
2027 Q2–Q3
P5planned
Platform
“The efficiency substrate other agents build on.”
2027 Q4+
The wedge is the cost curve. The precondition is auditable determinism.
The compounding asset is the accumulated, verified memory itself.