Skip to content
/
Phase B Batch 11

Batch 11 — Memory + Context Engineering

Batch 11 — Memory + Context Engineering

Roster (10)

slug stars distribution cli_binary local_ui orchestration multi_model tier
cline-mcp-memory-bank 60 mcp-server no no none no A
cline-memory-bank-nickbaumann 581 methodology-doc no no none no A
cline-memory-bank-official unknown methodology-doc no no none no C (docs URL)
cccmemory-mcp 26 npm-package yes (cccmemory) no none no A
ccmemory-plain 0 claude-plugin no no none no A (delta only)
claude-mem 78,248 npm-package yes (claude-mem) yes (port 37777) none yes (compression model) A
cc-memory 2 claude-plugin yes (mem.py) yes (Tkinter + browser) sequential yes (Haiku/Ollama) A
cc-simple-memory 0 claude-plugin yes (claude-memory) no none yes (Haiku/Sonnet/Opus) A
beads-yegge 24,094 cli-tool yes (bd, 80+ cmds) no task-decomposition-tree no A
beadsviewer 0 standalone-repo no yes (port 3000) none no A

Intra-batch Patterns

All 10 frameworks in this batch are variations on the same problem — AI coding agents forget everything between sessions — but they solve it at radically different levels of the stack. The file-based cluster (cline-mcp-memory-bank, cline-memory-bank-nickbaumann/official, ccmemory-plain) relies entirely on discipline and prompt engineering; the database cluster (cccmemory-mcp, claude-mem, cc-memory) uses SQLite or vectors for search; and beads-yegge uniquely uses a version-controlled SQL database (Dolt) reframing the problem as task-graph state rather than episodic memory. The two clear outliers are beadsviewer (pure UI companion, no agent features) and the cline-memory-bank-official entry (docs URL only). The most architecturally sophisticated entry is cc-memory (v2.2) with its anti-patch upsert_smart, forced PROGRESS.md handoff via system-reminder, PreCompact hook, and plan-refiner/plan-guardian subagents — despite having only 2 stars.

Most Interesting Finds

  1. beads-yegge — Fundamentally different framing: "memory" as a dependency-aware task graph in a Dolt (version-controlled SQL) database, not as session facts. The bd remember + bd prime pattern is the simplest possible cross-session injection mechanism in this batch, yet Dolt underneath enables multi-agent, multi-machine, multi-branch coordination that no other memory system supports. The "landing the plane" iron law in AGENT_INSTRUCTIONS.md is one of the most detailed and honest agent workflow contracts in the entire corpus.

  2. cc-memory (skymanbp) — Despite 2 GitHub stars, this is the most technically complete memory plugin in the batch: anti-patch upsert_smart (MERGE/SUPERSEDE/INSERT via trigram-Jaccard), forced PROGRESS.md read via <system-reminder> tag, PreCompact hook for compaction-boundary extraction, plan lifecycle with two subagents (plan-refiner, plan-guardian), FTS5 search, Tkinter + web UI. The gap between adoption (2 stars) and engineering quality is striking.

Items Written as Tier C

  • cline-memory-bank-official — Documentation URL (docs.cline.bot), not a code repository. Only 00-summary.md + METRICS.yaml written. Status: tier-c-evidence-incomplete.

Cross-References Discovered

  • ccmemory-plain is an explicit derivative of patrickkidd/ccmemory (seed) — same author, same plugin structure, Neo4j removed in favor of file-based hierarchy. Written as 4-file delta report in spec-system/ccmemory-plain-delta/.
  • cline-memory-bank-nickbaumann is the historical origin of the Cline Memory Bank pattern that was later canonicalized at docs.cline.bot (cline-memory-bank-official).
  • cline-mcp-memory-bank is explicitly inspired by roo-code-memory-bank (GreatScottyMac), which is not in this batch.
  • beadsviewer is a community companion to beads-yegge — reads data via bd --json CLI with no Dolt dependency.
  • cc-memory shares philosophical lineage with ccmemory (seed) in hook architecture but is an independent implementation (not a fork).
  • cccmemory-mcp was previously named claude-conversation-memory-mcp (renamed at v1.8.0) — confirm with the xiaolai/cccmemory slug.
← Batch 10
Batch 12 →