Skip to content

Proxy & ACT-R Memory

Six escalations to the proxy during this session. All under when_unsure policy. The escalation skill picked between answering autonomously and consulting the human at each one. The breakdown:

Conv id (suffix) Caller's question Human turns Terminal Mode
13edd7d7… INTENT review 0 RESPONSE: "NOT APPROVED" Autonomous
5fe9f71d… revised INTENT review 3 RESPONSE: "NOT APPROVED. Two changes required." Collaborative
b348d13c… revised INTENT review 2 RESPONSE: "NOT APPROVED — two targeted revisions required." Collaborative
ac1ea7b0… INTENT ratification 1 RESPONSE: "Both revisions are confirmed on disk." Collaborative
27d47559… PLAN review 1 RESPONSE: "Approve the plan with one redirect." Collaborative
3cb9ff07… WORK ratification 2 RESPONSE: "Ratified — proceed to APPROVED_WORK." Collaborative

This is the §7 escalation model operating end to end: one molecule (Proxy(t, x, π=when_unsure, h)) per call, the proxy non-deterministically choosing between [respond] and [escalate], every escalation eventually reaching [respond].

The autonomous escalation in detail

Of the six, the first one (INTENT review) is the cleanest demonstration of the autonomous path. The proxy ran a multi-step diligence pass without ever opening a dialog with the human:

  1. Loaded the escalation skill with argument when_unsure.
  2. Read collaborate.md (the per-policy escalation rail, since when_unsure routes there).
  3. Walked the worktree (which is a clone of the caller's worktree, per §7's "the proxy is a real agent").
  4. Read INTENT.md, IDEA.md, manuscript/architecture.md, and the per-chapter research artifacts that the prior session left behind.
  5. Read proxy.md — its own ACT-R memory bank.
  6. Decided, verbatim: "I'll respond from memory."
  7. Issued one RESPONSE JSON: "NOT APPROVED. Direct answers to your three questions..."

Zero dialog turns with the human. The decision to answer autonomously was the proxy's own — exercising the when_unsure policy's grant of judgment.

The proxy's reply was substantive: it identified that the INTENT under review "is not approvable as-is — it is missing the specific learnings from the prior attempt that was withdrawn at Pass 7 for executing the wrong intent." That specificity comes from the proxy having read both the current INTENT and the prior session's residual architecture artifacts. The §7 spec consequence — "an answer under never [or any policy where the proxy answers autonomously] is not a fiction; the proxy genuinely has the data and the memory to answer" — is observable here in the trace.

ACT-R memory in active use

The project's proxy memory at joke-book/proxy.md accumulated across sessions:

  • 17 ACT-R chunks existed before this session (earliest: 2026-04-23, latest pre-session: 2026-05-01).
  • Each chunk has structured frontmatter: id, type (procedural / institutional), domain (task / team), importance (0.5–0.95), reinforcement_count, last_reinforced.
  • A representative chunk:
## [2026-04-23] Observation
Category: Preferences
Signal: "Don't over think this."
Implication: When the human gives straightforward, direct instructions, execute
them without unnecessary analysis, elaboration, or suggestions for complexity.
Avoid over-design and premature consideration of edge cases or alternatives
when the task is clear and simple.

This is the Human Proxy system's ACT-R declarative memory adapted for gate decisions: structured chunks with activation-relevant metadata. The proxy reads these at every escalation as part of its diligence rail.

The consolidation log

This session's worktree contains .proxy-consolidation-log.jsonl recording 45 consolidation decisions:

Decision Count Meaning
ADD 38 New observation written as a fresh chunk
SKIP 7 Candidate observation merged into an existing chunk (or rejected)

The consolidation pass runs over candidate observations from the session and decides, for each pair, whether to ADD a new chunk or SKIP (because an equivalent chunk exists, or the observation is below threshold). This is the contradiction detection / consolidation pass that keeps the memory bank from drowning in duplicates.

38 new chunks added during one session is significant. The memory bank materially grew during this run — the proxy that runs the next session on this project starts from a richer memory than the proxy that started this one.

Escalation as a structural primitive

The six escalations weren't ad-hoc. Each was triggered by an AskQuestion from the joke-book-lead during its skill execution:

  • INTENT skill's ASSERT step (3 escalations to converge on approval, with revision rounds in between)
  • PLAN skill's ASSERT step (1 escalation; approved with one redirect)
  • EXECUTE skill's ASSERT step at end (2 escalations: one for an interim ratification, one for final WORK ratification)

The Proxy(t, x, when_unsure, h) molecule's behavior matched §7 across all six. Five of six routed through [escalate] at least once before terminating with [respond]. The autonomous case fired [respond] directly without ever entering [escalate].

The consolidated answers from the human's dialog in collaborative escalations were not just relayed — they were rewritten into the proxy's voice with ACT-R-grounded context. From the WORK ratification (verbatim from the bus): "Ratified — proceed to APPROVED_WORK and close. Darrell reviewed and signed off. Verbatim: 'it looks good.' Context on the ratification path: I walked the worktree, read Ch1 and Ch7 in full, sampled the reader-experience…"

The proxy carried both the human's literal approval and its own grounded confirmation that the work matched what was being approved. That second half is the ACT-R memory and the worktree-clone diligence rail composed.