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

Landscape

A survey of related projects in the AI agent memory and knowledge management space. Understanding what exists helps identify where kb-mcp fits, what patterns to adopt, and where to differentiate.

Quick Comparison

ProjectTypeLanguageSearchMCPWriteLocal/Cloud
kb-mcpKnowledge base serverRustBM25 + vector hybridstdioYesLocal
hipocampusAgent memory harnessJavaScriptBM25 + vector (qmd)NoYesLocal
obsidian-web-mcpRemote vault accessPythonripgrep full-textHTTPYesRemote
mengramCloud memory servicePython + JSSemantic (cloud)YesYesCloud
hmemHierarchical memoryTypeScriptTree traversalstdioYesLocal
mnemexSemantic code searchTypeScriptBM25 + vector (LanceDB)stdioNoLocal
cqAgent knowledge commonsPython + TSFTS5 + domain tagsstdioYesLocal + Team API
prism-mcpAgent session memoryTypeScriptFTS5 + sqlite-vec + TurboQuantstdioYesLocal or Cloud

Source Code Metrics

Measured with tokei. Source-only — excludes eval/benchmark data, web frontends, generated files, and node_modules.

ProjectLanguageFilesCode LinesCommentsTotal Lines
kb-mcpRust171,696362,001
hipocampusJavaScript373096954
obsidian-web-mcpPython181,469431,895
mengramPython (core)4822,0971,00725,907
hmemTypeScript106,5695777,617
mnemexTypeScript + TSX (src/)38886,91920,113120,031
cqPython + TypeScript8110,1291,66714,378
prism-mcpTypeScript + SQL12128,3678,58740,977

Notes: mengram includes cloud backend, SDKs, and integrations in its Python source. mnemex src/ includes the core engine, CLI, and MCP server — the full repo (329K lines) also contains eval benchmarks (91K), AI docs (15K), landing page, and VS Code extension.

Documentation Metrics

Markdown files as a proxy for documentation investment.

Project.md FilesLinesDoc-to-Code Ratio
kb-mcp524,8172.8x
hipocampus161,6572.3x
obsidian-web-mcp11980.1x
mengram171,8710.1x
hmem112,7100.4x
mnemex7423,8080.3x
cq122,2800.2x
prism-mcp91,4660.05x

Observations:

  • kb-mcp has the highest doc-to-code ratio (2.8x) — more lines of documentation than source code. This reflects the project’s dual role as both a tool and a documented reference implementation.
  • hipocampus also documents heavily (2.3x) — its memory architecture and protocol are extensively explained in markdown.
  • obsidian-web-mcp has minimal docs (single README) despite a substantial codebase — the code is the documentation.
  • mnemex has extensive docs (74 files, 24K lines) but the ratio is low because the codebase is so large.

Where kb-mcp Fits

kb-mcp occupies a specific niche: local, Rust-native, zero-infrastructure knowledge base server for curated markdown. It’s not trying to be agent session memory (hipocampus, hmem), a cloud service (mengram), or a code search tool (mnemex).

The closest overlap is with obsidian-web-mcp (both serve markdown vaults via MCP), but they differ on transport (local stdio vs remote HTTP) and search quality (BM25-ranked vs ripgrep grep).

The memory-focused projects (hipocampus, mengram, hmem, prism-mcp) are complementary rather than competitive — they manage agent session memory, while kb-mcp serves reference knowledge. An agent could use hmem for working memory and kb-mcp for its knowledge base.

cq represents a third category: collective agent learning. Where kb-mcp serves curated human-authored knowledge and memory projects persist agent state, cq captures wisdom that emerges from agent sessions and shares it across agents. All three categories coexist naturally.