LLM Memory & Knowledge Persistence¶
Patterns and architectures for organizing persistent knowledge when working with LLM agents. Not about building RAG systems (see [[llm-agents]]) - about how agents remember, forget, and manage knowledge across sessions.
Memory Architecture¶
- memory architectures - Flat files, hierarchical graphs, vector stores, knowledge graphs, hybrid approaches
- verbatim vs extraction - Why raw text beats LLM extraction (96.6% vs 85%), when to use each
Context & Window Management¶
- context window management - Layered loading (L0-L3), token budgets, compaction, re-injection patterns
Knowledge Organization¶
- knowledge base as memory - Raw->wiki->schema pipeline, plain markdown vs RAG, ingest/query/lint operations
- session persistence - Handoff files, memory files, state files, journal patterns
Temporal & Retrieval¶
- temporal memory - Validity periods, staleness detection, time-decay, contradiction resolution
- memory retrieval patterns - Index navigation, BM25, vector, hybrid search, reranking, cost comparison
Memory Lifecycle¶
- forgetting strategies - Compaction, archival, TTL, relevance scoring, memory size management