Build Log

Context Vault — Fixture Implementation

August 15, 2026

Built the real three-table schema, ingestion importer, stub compiler, and promotion pipeline against synthetic fixture data — with real, working enforcement of the candidate/canonical boundary and two genuine bugs the tests caught before they could reach real personal data.

ExperimentCodeAI MemorySQLite

files written (schema, ingestion, compiler, promotion, CLI, tests, fixture)

tests, all passing

real bugs found and fixed by the tests

real personal data touched

Timeline

Start

Export will take days — what to build meanwhile

Chose: build ingestion/schema/tests against synthetic fixture data now, so the real export is a drop-in swap later

Check

First-Draft Commitment Checkpoint asked and answered skip

Prior session's architecture.md already specified the three-table split and enforcement test requirement in enough detail — this session implements an already-reviewed design, not a new one

Build

Schema, ingestion, stub compiler, promotion, inspection CLI, fixture, tests

node:sqlite (Node 24 built-in), zero new dependencies

Spin

Batched Write tool calls silently failed to persist 6 of 9 files

Write tool reported success for every call; only the last file in a run of consecutive Write calls actually landed on disk each time

Verify

Ran the real test suite — found and fixed 2 real bugs

Stale compile-time target resolution, and a foreign-key insertion-order bug in the supersession logic

What shipped

experiments/context-vault/package.json — type:module, zero dependencies, node:sqlite (built-in, Node >=22.5.0)

experiments/context-vault/src/memory/schema.mjs — the three-table SQLite schema (raw_evidence/candidate_memories/canonical_memories) with FTS5 indexes, plus openCompilerConnection/openPromotionConnection/openReadConnection as restricted JS wrappers that structurally enforce the write-grant boundary from architecture.md

experiments/context-vault/src/ingestion/import-export.mjs — idempotent ChatGPT-export importer, quarantines malformed input into a warnings[]/fatal_errors[] report instead of silently dropping or crashing

experiments/context-vault/src/memory/compile.mjs — a rule-based STUB compiler (deliberately not an LLM call) proving the candidate-generation pipeline shape, enforcing the asymmetric evidence-weighting rule (assistant-only evidence can never produce origin=explicit_user_statement)

experiments/context-vault/src/memory/promote.mjs — the only path from candidate to canonical memory; auto-promotes explicit_user_statement candidates only, leaves system_inference candidates for human review

experiments/context-vault/src/api/inspect-cli.mjs — inspection CLI (list/memory/timeline/pending commands) per memory-model.md §26

experiments/context-vault/tests/fixtures/sample-export.json — hand-authored synthetic ChatGPT export matching the real mapping-tree format, deliberately including an assistant-authored claim about the user and a stated-then-reversed retirement goal

experiments/context-vault/tests/enforce-read-only.test.mjs — the Council Engineer's specifically-required enforcement test: mechanically enumerates exposed connection methods and fails if any write-shaped method appears on the read-only or compiler connections

experiments/context-vault/tests/memory-model-behavior.test.mjs — behavioral tests: idempotent ingestion, asymmetric evidence weighting, supersession-not-overwrite, provenance completeness, human-review gating

“the tests write themselves as soon as the fixture forces the hard cases to actually happen”
— Working note from this session — the Landon-soccer and retirement-reversal fixture cases weren't incidental, they were chosen specifically to force the asymmetric-evidence and supersession bugs into the open