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
| Project | Type | Language | Search | MCP | Write | Local/Cloud |
|---|---|---|---|---|---|---|
| kb-mcp | Knowledge base server | Rust | BM25 + vector hybrid | stdio | Yes | Local |
| hipocampus | Agent memory harness | JavaScript | BM25 + vector (qmd) | No | Yes | Local |
| obsidian-web-mcp | Remote vault access | Python | ripgrep full-text | HTTP | Yes | Remote |
| mengram | Cloud memory service | Python + JS | Semantic (cloud) | Yes | Yes | Cloud |
| hmem | Hierarchical memory | TypeScript | Tree traversal | stdio | Yes | Local |
| mnemex | Semantic code search | TypeScript | BM25 + vector (LanceDB) | stdio | No | Local |
| cq | Agent knowledge commons | Python + TS | FTS5 + domain tags | stdio | Yes | Local + Team API |
| prism-mcp | Agent session memory | TypeScript | FTS5 + sqlite-vec + TurboQuant | stdio | Yes | Local or Cloud |
Source Code Metrics
Measured with tokei. Source-only — excludes eval/benchmark data, web frontends, generated files, and node_modules.
| Project | Language | Files | Code Lines | Comments | Total Lines |
|---|---|---|---|---|---|
| kb-mcp | Rust | 17 | 1,696 | 36 | 2,001 |
| hipocampus | JavaScript | 3 | 730 | 96 | 954 |
| obsidian-web-mcp | Python | 18 | 1,469 | 43 | 1,895 |
| mengram | Python (core) | 48 | 22,097 | 1,007 | 25,907 |
| hmem | TypeScript | 10 | 6,569 | 577 | 7,617 |
| mnemex | TypeScript + TSX (src/) | 388 | 86,919 | 20,113 | 120,031 |
| cq | Python + TypeScript | 81 | 10,129 | 1,667 | 14,378 |
| prism-mcp | TypeScript + SQL | 121 | 28,367 | 8,587 | 40,977 |
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 Files | Lines | Doc-to-Code Ratio |
|---|---|---|---|
| kb-mcp | 52 | 4,817 | 2.8x |
| hipocampus | 16 | 1,657 | 2.3x |
| obsidian-web-mcp | 1 | 198 | 0.1x |
| mengram | 17 | 1,871 | 0.1x |
| hmem | 11 | 2,710 | 0.4x |
| mnemex | 74 | 23,808 | 0.3x |
| cq | 12 | 2,280 | 0.2x |
| prism-mcp | 9 | 1,466 | 0.05x |
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.