docs(memory): expand memory chapter with long-term memory, patterns and pipelines

[Thiago Rotta (rottathiago)]

What

Evolves the memory section from an introduction stub plus a storage-focused STM page into a complete chapter. All existing STM content was preserved and re-framed, not rewritten.

Pages

Page	Status	Content
Memory.md	expanded	Memory types (STM, LTM, working memory), cognitive sub-types (semantic / episodic / procedural), memory vs. knowledge sources, key principles, chapter map
Memory-Architecture-Patterns.md	new	Storage strategies, retrieval patterns, vendor-neutral pattern archetypes, memory scoping, trade-offs (incl. the fading memory problem), token economics, metrics
Short-Term-Memory.md	extended	Existing design approaches and retention kept; adds context window management, working memory assembly, session end handling, multi-channel considerations
Long-Term-Memory.md	new	What to persist, memory entry data model, lifecycle, relevance scoring and decay, scope decision framework, governance, security risks, operations at scale, phased roadmap
Memory-Pipelines.md	new	Mermaid diagrams for the retrieval, session end, and STM → LTM promotion pipelines

Also: SUMMARY.md navigation updated (the empty long-term memory entry is now filled) and the Mem0 long-term memory benchmark paper added to docs/References.md.

Highlights for reviewers

• Relevance scoring and decay (Long-Term-Memory.md): score composed from retrieval frequency, recency, importance and confidence; last_retrieved_at / retrieval_count write-back on the read path; use it or lose it decay; and tier demotion (hot → warm → cold → archived) instead of hard deletion, with retrieval promoting entries back up.
• Pipelines (Memory-Pipelines.md): the three flows are authored in Mermaid, each with a stage table and design notes, plus an overview showing how the read path, session-end path and promotion path connect.
• Market implementations are abstracted into neutral pattern archetypes (auto-injected layered, on-demand tool-based, permission-trimmed retrieval, extract-and-update, self-managed tiered) rather than named vendors, with a hybrid recommendation for enterprises.
• Technology-agnostic first, with Azure services named only as concrete examples.
• The adoption roadmap intentionally describes phases by objective and outcome, without time estimates.

Validation

• markdownlint clean across docs/**/*.md
• prettier --check clean (80 cols, proseWrap: always)
• python scripts/update_references.py reports no new links
• All 13 Mermaid blocks verified by rendering them with mmdc
• All internal links and heading anchors resolve

[Fernando de Oliveira (fedeoliv)]

Really nice content, thank you Thiago Rotta (@rottathiago)!