---
source_hash: 4df43be456b759191b327b1e533798a2b6c614670e5489ed68301eb64fc1b9bb
post: /posts/memory/generative-memories
shape: doctrinal
---

# Context Pack: Generative Memories

Agent reference for the rewritten essay. Only the quoted excerpts below are verbatim. The release report describes the original August 2026 publication, not newly verified live infrastructure.

## Verbatim excerpts

> The build is the instrument. The memory is the product.

> The build is real as an experiment. The memory is counterfactual: this was a road we didn't take in the live product, and the note must never imply otherwise.

> A previously failed key remains eligible for its single later retry.

> A cap pause was sticky. Clearing it required a human action, not an automatic retry.

## Purpose and design

Build one rejected approach as a throwaway ghost and distil transferable learning into a labelled synthetic memory. A rejected design summary alone is not new experimental experience. The live product still follows its selected design.

Six discovery legs, three drafts and three judges produced three rejected proposals with salvage. The hybrid took risk ordering from risk-first, a single-file pipeline and local expansion from MVP-first, configuration and provenance boundaries from portability-first, and transferable insight as the goal.

Five milestones run in order: inspect the live system; prove synthetic provenance through recall; probe a headless executor with recorded exit code/logs; complete one real ghost and have Ari score it; then enable the timer. Recon found a dead distiller stub, file-scan ingestion through a path-to-layer map, no recall layer filter, and rejected-design digests already entering as raw fact.

## Provenance contract

Use a dedicated synthetic layer; reject a sink also claimed by a non-synthetic ingestion rule; mark frontmatter and use a non-colliding filename prefix; put a fixed warning on body line one and prove it survives recalled chunks. A banner's presence in the source is not sufficient evidence that retrieval exposes it.

The counterfactual card's banner is `SYNTHETIC COUNTERFACTUAL // this build never shipped`. Inspect actual recalled output rather than relying on frontmatter surviving.

## Pipeline and limits

Intake prefers the oldest unledgered eligible kill verdict and handles one approach per run. Recognise multiple digest formats and deduplicate by slug, approach ID and normalised sketch. Failed keys may retry once on a later run, then park and surface in status. Salvage intake is deferred.

Expand the roughly 250-word sketch locally, retaining its rejection reasoning. Give the mini-spec to one headless builder in the ghost directory. Check cost before the call; enforce a per-run cap, timeout backstop and append-only spend ledger.

Distil from the spec, transcript and rejection reasoning. Validate required sections, substance and filler before emitting only to the synthetic sink. The configured cadence was every three days, about ten runs a month. A cap breach pauses further work until a human clears it. Status and dry-run are manual surfaces; the timer drives normal operation.

The six required memory sections, in order, are: What surprised; Technique emerged; Transfers to; Kill verdict confirmed/refuted; Build shape; Open questions.

## Four failures and first result

1. Builder narrated but wrote no files: require file-writing acceptance and fail an empty build.
2. Terminal control codes polluted distillation: preserve the raw log, strip and normalise the input.
3. Brittle validation: check required sections and minimum substance length rather than superficial formatting.
4. Reasoning-style preamble leaked into the note: strip recognised framing and revalidate.

The fourth run emitted a memory, scored by Ari before trusting the timer. It tested the rejected work-not-people template and built a validator. The How axis admitted persona language; Linguistic Linting using negative constraints helped enforce task definitions. The original kill was confirmed for the template and salvaged for a validator. Transfer: orchestration can change model identities without rewriting work around personas.

## Release and exclusions

The original release reported a tested single-file CLI, synthetic provenance checked end to end, the first scored ghost, a three-day timer and a per-run cap. Remaining work included distillation quality, a stronger stamp route, a morning-panel failure line and salvage intake. A paid/stronger model was an available quality dial, not automatic extra spending.

V1 deferred multiple executors beyond a configuration seam, salvage intake, brain-side recall filters and remote pushes. Automatic ghost deletion was rejected permanently; status may report disk use and Ari decides what to prune. Do not recast that boundary as a deferred cleanup feature. Human scoring gates the first result, not every later emission.