| name | 0.4.3-understand-long_term_memory |
| description | [0.4.3] Information persisted across sessions, tasks, or deployments; requires external storage and retrieval mechanis |
understand-long_term_memory
CALL NUMBER: llm_memory_and_context_engineering.long_term_memory : deep_long_term_memory(14)
DEFINITION: Information persisted across sessions, tasks, or deployments; requires external storage and retrieval mechanisms since it cannot fit in context_window
Invoke this skill to understand long_term_memory down to its primitives. The RELATIVE ROOT below is the least-fixed-point closure of everything it bundles from — the full import cone, grouped by the lib each prim comes from. Projected from a prover-typed KB (MAP/SWI-Prolog consistency gate): every reference below resolves.
THE RELATIVE ROOT (the import cone, by lib)
from deep_long_term_memory
- em_episode_marker (d2): Signal indicating episode boundaries: new session start, task boundary, or conversational turn demarcation.
- em_event (d2): A discrete unit of experience stored in episodic memory — a single interaction, task completion, or notable outcome.
- em_significance_rank (d2): Ordering structure by event importance or impact; enables retrieval by significance rather than recency.
- em_temporal_index (d2): Organizational structure mapping events to time ranges; enables retrieval by temporal proximity.
- em_experience_sequence (d2): An ordered collection of events forming a coherent experiential arc; the episodic unit of storage.
- em_salience_weight (d2): Numeric value indicating how memorable or significant an event is; influences consolidation priority and retrieval ranking.
- em_recency_decay (d2): Algorithm for reducing salience weights of older events over time; models forgetting in episodic memory.
- kg_inference (d2): Deriving new kg_triple entries from existing ones through rule application or traversal over the kg_ontology (e.g., transitivity of is_a)
- kg_traversal (d2): Navigating the knowledge_graph along kg_relationship edges to discover connected kg_entities; the fundamental operation enabling associative retrieval
- em_context_bundle (d3): The surrounding context of an event including user inputs, system responses, and environmental state at the time.
- em_interaction_record (d3): Structured log of the exchange pairs (prompt/response) comprising an event.
- em_outcome (d3): The result or consequence of an event — success, failure, or partial completion — recorded for outcome-based retrieval.
- em_state_snapshot (d3): Captured system state (context window contents, active variables, user profile) at the moment of an event.
- em_timestamp (d3): Temporal marker indicating when an event occurred; enables chronological ordering of episodic entries.
from llm_memory_and_context_engineering
- episodic_memory (d1): Memory storing discrete experiences or events — specific interactions,完成任务, or notable outcomes — organized chronologically or by significance
- knowledge_graph (d1): A graph-structured representation of entities and their relationships; used as a structured form of semantic_memory with traversable edges
- memory_audit (d1): Systematic review of stored memories for accuracy, relevance, and compliance; governance practice for long_term_memory systems
- memory_capacity (d1): Maximum storable items in a memory system; bounded by context_length for working_memory and storage size for long_term_memory
- memory_consolidation (d1): The process of transforming short-term memory representations into longer-term storage; converting ephemeral state to persistent records
- memory_invalidation (d1): Removing or updating outdated memories; necessary when world_model facts change and stored memories become incorrect
- memory_persistence (d1): How long memory survives; session memory (within conversation), user memory (across sessions), or permanent memory (until explicitly deleted).
- memory_poisoning (d1): Corruption of long_term_memory stores with false or harmful content; defense requires verification and source attribution
- memory_retrieval (d1): The act of fetching stored memory entries — by semantic similarity, keyword, metadata, or temporal proximity — to include in context
- semantic_memory (d1): Memory storing general world knowledge, facts, and concepts rather than specific personal experiences; queried via retrieval for grounding
- semantic_network (d1): Graph of concepts connected by typed edges; models can store and traverse this structure for relational reasoning
- vector_database (d1): Storage system indexing embeddings for similarity search; backbone of RAG. Examples: Pinecone, Weaviate, Chroma, FAISS. Enables retrieval from vast document collections.
- semantic_similarity (d2): Measure of meaning overlap between two text fragments in embedding_space; the basis for associative_memory retrieval
- approximate_nearest_neighbor (d2): Fast retrieval of roughly similar vectors accepting some error; enables scaling vector_database beyond exact search limits
- chunking_strategy (d2): How documents are split into retrieval units; choices affect recall/precision: fixed-size, semantic (by sentences/paragraphs), recursive, or agentic splitting.
- cosine_similarity (d2): A similarity measure between embedding vectors based on the cosine of the angle between them; invariant to vector magnitude
CONSUMERS (what needs this)
retrieval_augmented_generation, system_message
Projected from the llm memory and context engineering KB (438 concepts / 213 relations) — consistency-typed by MAP; the facet list after the colon IS the cross-lib dependency web.
(leaf — this is an actual skill.)
