Memory agent audit report and roadmap documentation

[jeremiepas]

Summary

Complete audit of Graphos as a memory agent platform, including current capabilities, gap analysis, target architecture, implementation roadmap, and technical specifications.

Documentation created

• docs/proposals/memory-agent/00-README.md — Overview + TL;DR
• docs/proposals/memory-agent/01-current-capabilities.md — What already works (70%)
• docs/proposals/memory-agent/02-gap-analysis.md — 8 gaps ranked by priority
• docs/proposals/memory-agent/03-architecture.md — Target architecture
• docs/proposals/memory-agent/04-implementation-roadmap.md — 5 phases, 13-18 days
• docs/proposals/memory-agent/05-technical-specifications.md — Detailed specs

Key Findings

• Graphos is ~70% of a memory agent already
• MCP server exposes select_context, add_conversation, conversation_history
• Critical gaps: mutable graph state, semantic search, LLM summarization
• Clean architecture (Domain/UseCase/Infrastructure) makes enhancements straightforward

Related Issues

• Make MCP server graph state mutable (TVar) #5 Mutable MCP state
• Cross-session graph persistence (memory/ → graph merge) #6 Cross-session persistence
• Add semantic search with embedding-based retrieval #7 Semantic search
• LLM-driven conversation summarization #8 LLM summarization
• Add temporal awareness to context selection #9 Temporal awareness
• Add incremental graph mutation via MCP (add_node, add_edge) #10 Incremental mutation
• Add bidirectional conversation-code reverse index #11 Bidirectional reverse index
• Implement DifferentialContext selection strategy #12 Differential context
• Performance optimization: select_context on large graphs #13 Performance optimization