27. Memory

Flux leaves a Seed; Archive preserves its return; Recall moves retained form into present work.

Decision

LychD adopts PostgreSQL Archive with pgvector-derived indexes behind LychD-owned ports. PostgreSQL owns authoritative records, lineage, policy, lifecycle, and derivations; frameworks may extract, embed, rank, or rerank but cannot silently own transaction, migration, curation, sharing, or training eligibility. Run/Step ledgers, traces, checkpoints, Personas, one-call context, and Soulforge corpora stay distinct and cite by stable ref.

Framework boundary and Memori

Memori is a maintained comparison source, not an installed LychD subsystem. Its explicit entity/process/session attribution, raw-conversation versus derived augmentation stages, idempotent conversation creation, bounded background writer queue, and drain operation are useful implementation patterns. A pinned source checkout may support design review; it grants no runtime or schema authority.

LychD does not adopt transparent LLM-client interception, retain every conversation, automatically extract claims, or inject recall into every prompt. Those conveniences cross Archive's admission, privacy, correction, and Context authority. A future LychD-owned CandidateArchivePort must admit attributed raw candidates and separately identified derivatives, preserve source lineage and anti-reingestion identity, and expose bounded processing state. The exact retry and stale-write rules remain design requirements. Embeddings, recall, curation, promotion, RAG injection, and training remain later consumers with their own evidence.

The Archive record

One immutable revision names stable record/revision, namespace owner/subject, kind/lifecycle, content or immutable ref, source/producer/observation/transformations, creation-observation-validity- expiry-supersession-contradiction-retention times, classification/sharing/deletion authority, quality/evaluator evidence, related Session/Run/Invocation/Pattern/Persona/artifact refs, and derived representations.

candidate → promoted → archived
    │           │
    └───────────┴→ revoked

candidate awaits adjudication; promoted is eligible only for declared recall; archived remains but is excluded from ordinary recall; revoked retains lawful tombstone/lineage while content/indexes are removed or quarantined. Promotion is bounded eligibility, never infallibility: Viparyaya and Pramāṇa both carry source and correction.

Namespaces and authority

Write and recall name namespace before retrieval. A namespace may be Principal, Persona, Composition, shared body, or published corpus, but id/policy are explicit. Current Sigils lack a stable entity_id: Archive must not invent Sigil.id or claim delivered cross-identity isolation. Default is no cross-namespace recall. Sharing identifies source, consumer, purpose, record/field classes, duration, onward disclosure, revocation. Similarity, organization, delegation, and model do not grant it. Authorize candidates before content or embeddings leave owner boundary.

Embeddings are derived data

A vector inherits source chunk Privacy Label and deletion lineage; opacity never authorizes remote embedding. Context owns labels and Security declassification. Each derived representation records source revision/chunk, embedder identity/revision/digest/configuration, dimension/distance, normalization/chunking/preprocessing, creation/status/quality receipt, and index generation. Incompatible spaces never compare. New embedder creates a new generation; old may remain policy- queryable during proved migration, then retires, never overwritten/mixed. pgvector makes one governed DB backup target, not a coherent snapshot of artifacts/models/services. Schemas, HNSW, lexical/hybrid indexes are deployment work, not current claim.

Ingestion

Ingestion is admitted workflow:

1. authorize source and memory purpose;
2. retain immutable source/provenance;
3. make candidate units with versioned parser/extractor;
4. classify privacy, owner, retention, sharing;
5. validate and retain uncertainty/contradiction as such;
6. write candidate revisions;
7. derive representations through admitted embedding capability;
8. offer eligible candidates to Curator.

Workers may partition/embed; Dispatcher picks eligible capability and Orchestrator readiness, but neither decides worth. Extractor facts/relations/summaries/preferences are attributed claims with source excerpt/ref. Authoritative record and derivation may commit separately: asynchronous vector work exposes honest index state and stays invisible to vector recall until complete generation; retry is idempotent by source revision and derivation spec.

Curation and sediment

Versioned Curator considers source quality, verification, correction, contradiction, use outcome, recency, expiry, and Riddle findings. Access, repetition, similarity, and praise are not truth; they can affect salience only by declared policy. Curator may promote, retain, archive, revoke and invalidate derivatives, or relate supersession/contradiction without rewriting history. Anchors have owner/review rule, not immortality. Batch curation stages revisions and never changes active agent context. Mirror, Context, and Riddle may use records but do not become Curator.

Recall

One contract serves Pattern context or authorized tool: bind caller/namespace/purpose/policy/query and result/token budget; authorize fields/classes; choose compatible lexical/vector/relational/hybrid plan; retrieve/rerank versionedly; enforce threshold/diversity/recency/contradiction; return bounded provenance/lifecycle/times/uncertainty; write receipt without hidden prompt as memory. Similarity is position, not truth probability; threshold miss is no admissible result, not absence from world. Context fits results and records omissions; model sees attributed prior, not instruction. Receipt pins policy and returned revisions; later correction supersedes future queries and preserves earlier Run influence.

Continuity and deletion

Committed rows survive process death; Reanimation needs explicit Pattern and compatible revisions, not restored thought or Persona. Deletion removes/quarantines controlled content/indexes, keeps minimum anti-reingestion tombstone, invalidates dependent recall/evaluation, and identifies exported or shared copies plus Soulforge descendants. It cannot erase influence in generated artifacts or weights; those need their own decisions.

Training boundary

Archive is structured Soulforge input, never corpus by default. Training selects exact revisions and independently admits privacy/license/dedup/split/holdout. Findings/feedback may nominate review, never automatic Karma/rank/weight change.

Rejected alternatives

A vector service as the source of truth

A derived store may scale retrieval, but cannot own content, provenance, policy, lifecycle, or restore.

A retrieval framework as the Archive

An add/search API cannot expose curation, deletion, audit, and training-admission records.

Automatic retention of every trace

It expands privacy exposure and feeds error back into context; only explicit ingestion makes candidates.

Consequences

Acceptance evidence

The current Partial claim covers only the candidate-admission contract and volatile adapter. Broader Partial acceptance requires one class proving authorized ingestion, provenance candidate, compatible embedding, namespace recall, threshold miss, correction, staged promotion, deletion/index cleanup, restoration, and reproducible receipt. State records delivery.