Skip to content
/
Phase B Batch 26

Batch 26 — Memory Secondary (Local-first, Episodic, KB Infra)

Batch 26 — Memory Secondary (Local-first, Episodic, KB Infra)

Roster (9)

slug stars distribution cli_binary local_ui orchestration multi_model tier
rlm-claude 35 mcp-server no no parallel-fan-out no A
nemp-memory 101 claude-plugin no no none no A
memoryagent 38 skill-pack no no none no A
nex-as-skill 44 npm-package yes (nex) no none no A
contextmcp unknown mcp-server unknown unknown none unknown C
agentskb unknown mcp-server unknown unknown none unknown C
context-engineering-handbook 9,009 methodology-doc no no none no A
context-space 810 standalone-repo no no none no A
context-mode 15,710 npm-package yes (context-mode) yes (web-dashboard, port 3000) none no A

Intra-batch Patterns

The nine frameworks cluster into four clear groups: (1) local-first plain-file memory (nemp-memory, memoryagent) — maximally simple, zero deps, zero cloud; (2) MCP-anchored persistent memory (rlm-claude, nex-as-skill) — richer search/retention but require server infrastructure; (3) context-window efficiency infrastructure (context-mode, context-space) — focus on what goes INTO the context, not what persists across sessions; and (4) educational/methodological content (context-engineering-handbook). A striking pattern: every framework except context-space targets session continuity after /compact or context loss as the primary use case — the MIT RLM paper (arXiv:2512.24601) appears to be the shared academic ancestor, cited explicitly by rlm-claude and indirectly referenced in the handbook. None of the frameworks in this batch enforces TDD or automates git operations — they are pure memory/context infrastructure, not development workflow frameworks. The star counts are bimodal: three frameworks have <45 stars (rlm-claude, nemp-memory, memoryagent) while two have >9,000 (context-engineering-handbook at 9k, context-mode at 15.7k).

Most Interesting Finds

context-mode (15,710 stars, HN #1): The "Think in Code" paradigm is the most novel agent architecture insight in the entire research corpus — agents should generate analysis scripts that return only results rather than read data into context, achieving 100x more context savings than any other technique. The SQLite+FTS5+BM25 session continuity + analytics web UI with 90 metrics make this the most technically sophisticated memory/context framework in the batch.

context-engineering-handbook (9,009 stars): The only pure methodology handbook in the catalog. Uniquely applies dynamical systems theory (attractor dynamics, semantic resonance) as engineering primitives — not as metaphor but as concrete design patterns. Conceptually upstream of every other framework in this batch.

Items Written as Tier C

  1. contextmcp — URL: contextmcp.ai. No public GitHub repository found. Website-only SaaS service with no inspectable source code.
  2. agentskb — URL: agentskb.com. No public GitHub repository found. Website-only SaaS service with no inspectable source code.

Cross-References Discovered

  • rlm-claude directly implements the MIT CSAIL RLM paper (arXiv:2512.24601); context-engineering-handbook cites the same paper as a key reference — both are implementations/theorizations of the same research lineage.
  • nemp-memory explicitly positions against ccmemory (the seed) and unnamed MCP-based memory systems as "too complex."
  • context-mode and rlm-claude both hook into PreCompact but take opposite philosophies: rlm-claude blocks compaction and requires a user save decision; context-mode automatically snapshots and restores via BM25.
  • nex-as-skill and context-space both aggregate external services (Slack, HubSpot, GitHub) but via different mechanisms: Nex as cloud knowledge graph, context-space as real-time MCP tool execution.
  • memoryagent and nemp-memory share the "plain files" philosophy but differ in complexity: memoryagent is 1 skill with 6 sub-commands; nemp-memory is 24 commands with hooks and agent identity tracking.