Keyboard shortcuts

Press ← or → to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

Roadmap

What’s been built, what’s next, and when each phase makes sense.

Completed

Standalone MCP Server (v2)

RON-configured binary with 10 MCP tools, full CLI parity, and zero hardcoded project-specific values. Open sourced at github.com/ttdonovan/kb-mcp.

Persistent Storage (memvid-core)

Replaced in-memory Tantivy with memvid-core .mv2 persistent files. Incremental reindex via blake3 content hashing. Smart markdown chunking. Crash-safe WAL.

Containerized Researcher Agent

ZeroClaw container with kb-mcp + DuckDuckGo web search. Writes research findings to vault/drafts/ for human review. IDENTITY.md/SOUL.md define agent personality and quality standards.

Hybrid Search (kb-mcp Phase 2)

BM25 + vector search via memvid-core vec feature. Local ONNX embeddings (BGE-small-en-v1.5), HNSW vector index in .mv2 files, RRF fusion. Opt-in via cargo build --features hybrid. Container supports just agent-build-hybrid.

Vault Intelligence Bundle

Three new MCP tools (kb_digest, kb_query, kb_export) plus transparent auto-reindex via directory mtime checks. Brings the tool count from 6 to 9. kb_write also gained optional directory and filename parameters for hierarchical collection structures.

Cargo Workspace Split

Reorganized from a single binary crate into a three-crate workspace: kb-core (shared library), kb-cli (binary kb), kb-mcp-server (binary kb-mcp). Eliminates CLI/MCP code duplication, guarantees behavioral parity through shared kb_core::format::* functions, and enables independent compilation with clean dependency isolation.

Up Next

Draft Reviewer Agent

A second container agent (or Claude Code sub-agent) that reviews drafts for quality, formatting consistency, source verification, and proper frontmatter — then promotes approved entries into the vault.

Why now: The researcher agent is producing drafts. Manual review is the bottleneck. Automating the quality gate completes the capture pipeline.

Scope:

  • Read drafts collection, check against vault conventions (SOUL.md standards)
  • Verify sources are reachable URLs
  • Ensure frontmatter has required fields (tags, created, updated, sources, target)
  • Promote approved drafts to the correct vault section
  • Flag issues for human attention rather than silently fixing

Heartbeat Scheduling

Add HEARTBEAT.md to the researcher agent for automated periodic research.

Why now: The researcher works well manually. Scheduling is a small addition that makes it run on autopilot for configured topics.

Scope:

  • HEARTBEAT.md defines research topics and frequency
  • ZeroClaw cron runs the research workflow on schedule
  • Digest report of what was added (file or notification)

Future

Knowledge Capture Tools (kb-mcp Phase 3)

Specialized write tools beyond free-form kb_write:

  • kb_capture_session — record debugging/coding sessions
  • kb_capture_fix — record bug fixes with symptom/cause/resolution
  • kb_classify — auto-tag unprocessed notes (type, tags, summary)

When: When agents are actively writing and would benefit from structured capture templates.

Gap Analyzer Agent

Reads the existing vault, identifies thin or missing topics, and feeds the researcher agent with specific research requests. The inward-facing complement to the outward-facing researcher.

When: When the vault is large enough that gaps aren’t obvious from browsing.

Knowledge Keeper Agent

Combines researcher + reviewer + gap analyzer into the full Knowledge Keeper pattern. Sweeps sessions, scores knowledge by usefulness, prunes stale entries. The most autonomous form of vault curation.

When: When all three component agents are proven individually.

HTTP Daemon Mode (kb-mcp Phase 4)

Add HTTP transport alongside stdio. Long-lived server eliminates MCP cold starts and enables network-based access (no volume mount needed for container agents).

When: When cold start latency is a problem (likely after hybrid search adds ONNX model loading).

Cross-Agent Knowledge Sharing (kb-mcp Phase 5)

Multiple projects share knowledge through federated .mv2 files. Agents in different repos contribute to and query from shared collections.

When: When multiple projects are actively using kb-mcp and would benefit from shared context.