[EPIC][CORE-D] Semantic projection + corpus ingestion: compiler adapter, projection generations, structured adapters, resumable backfill, and artifact lineage #34

New issue

Open

opened 2026-09-08 01:57:02 +00:00 by nsaspy · 0 comments

nsaspy commented

2026-09-08 01:57:02 +00:00

Owner

Parent: #11
Refines: #5 #8
Upstream: prolog-rlm#388 #393 #396

Goal

Implement the ingestion side of Symbolic Memory: exact durable sources become independently versioned semantic projections, and large corpora can be compiled/resumed without replaying everything.

Own:

compiler adapter over Prolog-RLM public APIs;
projection lifecycle absent|pending|ready|failed|stale|blocked;
source/compiler/schema/config fingerprints;
exact source-span/semantic-record lineage;
structured deterministic adapters for LLM logs, issue/PR streams, Org/Markdown, JSON/JSONL, web-result records and traces;
stable segmentation/work IDs;
resumable jobs/checkpoints/cancellation;
stage-level cache/artifact retention where required for replay;
backfill of existing source memories missing projections;
idempotent re-run and selective invalidation when source/compiler/schema changes.

The semantic compiler itself remains in Prolog-RLM. Symbolic Memory owns durable corpus/job/projection lifecycle.

Hard invariants

source commit success != projection success
projection generation != source version
same source/config retry must not duplicate current semantic records
changed source/compiler/schema creates new lineage
partial job failure != coherent publication
model requery != exact historical replay unless artifact retained

Acceptance

Ingest and resume mixed structured/unstructured corpus fixtures; kill/restart mid-run; prove no duplicate completed projection work; preserve exact source/span/compiler lineage; and backfill old memories without changing their original source IDs.

Parent: #11 Refines: #5 #8 Upstream: prolog-rlm#388 #393 #396 ## Goal Implement the ingestion side of Symbolic Memory: exact durable sources become independently versioned semantic projections, and large corpora can be compiled/resumed without replaying everything. Own: - compiler adapter over Prolog-RLM public APIs; - projection lifecycle `absent|pending|ready|failed|stale|blocked`; - source/compiler/schema/config fingerprints; - exact source-span/semantic-record lineage; - structured deterministic adapters for LLM logs, issue/PR streams, Org/Markdown, JSON/JSONL, web-result records and traces; - stable segmentation/work IDs; - resumable jobs/checkpoints/cancellation; - stage-level cache/artifact retention where required for replay; - backfill of existing source memories missing projections; - idempotent re-run and selective invalidation when source/compiler/schema changes. The semantic compiler itself remains in Prolog-RLM. Symbolic Memory owns durable corpus/job/projection lifecycle. ## Hard invariants ```text source commit success != projection success projection generation != source version same source/config retry must not duplicate current semantic records changed source/compiler/schema creates new lineage partial job failure != coherent publication model requery != exact historical replay unless artifact retained ``` ## Acceptance Ingest and resume mixed structured/unstructured corpus fixtures; kill/restart mid-run; prove no duplicate completed projection work; preserve exact source/span/compiler lineage; and backfill old memories without changing their original source IDs.

nsaspy referenced this issue

2026-09-08 01:57:39 +00:00

[EPIC][CORE] Library-first Symbolic Memory v1: durable semantic world model, learning history, federation, and zero-model recall #38