Memory
# Spectron Principles (Reference) Source: https://surrealdb.com/platform/spectron ## 8 Pillars 1. **Authoritative** — Curated artefacts (documents, manuals, policies). Higher default trust. 2. **Experiential** — Conversational input. Lower default trust. 3. **Reconciliation** — One function for all writes. Supersession + uncertainty path. 4. **Elaboration** — Background sweep finds entities sharing context but no relation. 5. **Reflection** — On-demand synthesis. Persisted as new facts with lower trust. 6. **Consolidation** — Belief crystallization. Create, update (delta recorded), or mark superseded. 7. **Calibration** — Source trust + reconciler confidence, both queryable. Low-confidence never auto-supersedes high-confidence. 8. **Collective** — Shared, reconcilable memory across instances. ## 6 Memory Categories 1. Episodic — Raw conversational record (source of truth) 2. Identity — Durable facts about principal (long retention, low decay) 3. Knowledge — Things learned/shared (medium retention) 4. Context — Current working set (short retention, replaced rapidly) 5. Instructions — Behavioral memory (applied at prompt-assembly time) 6. Uncertainty — Explicit "we don't know" rows ## 5 Coherence Dimensions 1. Semantic — Embedding-based recall 2. Lexical — BM25 over surface text 3. Relational — Entity and relation graph (1-2 hops) 4. Temporal — Three clocks (system, known, valid) 5. Spatial — Geographic recall (first-class) ## Hybrid Retrieval (8 signals) 1. Vector recall (dense embeddings) 2. Lexical BM25 3. Graph traversal (1-2 hops) 4. Keyword bridges (RAKE + PMI) 5. Document links (section embeddings, doc-to-doc edges) 6. Personalised PageRank 7. Geographic recall 8. Trace-derived features ## Tiered Queries - Tier 1: Direct lookup (sub-ms, no embeddings/LLM) - Tier 2: Response reuse (cached, entity-aware invalidation) - Tier 3: Hybrid retrieval (vector+BM25+graph+keywords, LLM synthesis) - Tier 4: Full-context fallback (broader sweep, deeper traversal) ## Key Patterns - Supersede, don't delete (valid_from/valid_until, history preserved) - Provenance as data on every fact (kind, ref, trust, lexical_span) - Three clocks: system_time, known_time, valid_time - Content-addressed (Blake3 hash, re-upload is no-op) - Same reconciler for all write paths - Traces are memory (first-class graph nodes) - Uncertainty is explicit (confidence below floor → emit uncertainty row) - forget is explicit (not natural aging, with --purge option)
Tags: spectron, reference, memory-model, retrieval, project:intelligentrag, kind:semantic — Source: claude — 2026-07-16 18:10:38 UTC