Knowledge & MCP Tools

lm-assist automatically builds a searchable knowledge base from your Claude Code sessions and makes it available through MCP tools and context injection.

How Knowledge Is Built

When Claude Code runs Explore subagents (research tasks), lm-assist captures the results and extracts reusable facts:

Claude Code Session
       |
       v
  Explore subagent returns findings
       |
       v
  lm-assist extracts structured knowledge
       |
       +----> Vector index (semantic search via LanceDB)
       +----> BM25 index (keyword search)
       +----> Knowledge store (full content)

Each knowledge entry contains:

• Title -- what was explored
• Facts -- specific, reusable findings (code patterns, API details, file locations)
• Source -- which session and project it came from
• Timestamps -- when it was created and last updated

Knowledge Entry Example

K025: Explore user authentication flow

Facts:
1. Auth middleware at gateway-type1/lib/middleware/auth.ts
2. JWT tokens stored in httpOnly cookies
3. OAuth callback at /v1/auth/callback/{provider}
4. Supported providers: Google, GitHub, Microsoft
5. Guest access creates temporary user with 24h expiry

Source: Session abc123 | Project: /home/user/LangMartDesign
Updated: 2026-02-20

MCP Tools

lm-assist registers three MCP tools in Claude Code. These are also available in any MCP-compatible IDE (VS Code, Cursor, Windsurf, etc.).

search

Search the knowledge base using natural language, keywords, or file paths.

search(query, scope?, project?, type?, limit?, offset?)

Example usage in Claude Code:

When Claude needs to understand a part of your codebase, instead of re-exploring files, it calls:

search("authentication middleware implementation")

And gets back relevant knowledge entries from previous sessions.

detail

Get the full content of a knowledge entry by ID.

detail(id, section?, limit?, offset?)

feedback

Provide quality feedback on knowledge entries. This improves future search results.

feedback(id, type, content)

Example: If a knowledge entry references an old file path:

feedback("K025.3", "outdated", "File moved from lib/auth.ts to middleware/auth.ts")

Context Injection Hook

The context injection hook is the most impactful feature. It fires automatically before every Claude Code prompt:

1. Prompt received -- the user types a message in Claude Code
2. Hook fires -- lm-assist searches the knowledge base for relevant entries
3. Results injected -- top matches are added as context to the system prompt
4. Claude responds -- with knowledge from previous sessions already loaded

What Gets Injected

The hook injects a compact summary with:

• Knowledge entry IDs and titles
• Key facts relevant to the current prompt
• Suggestions to use detail(id) for deeper exploration

Typical injection size: 2-5 entries, ~3k tokens.

Viewing Injection Logs

/assist-logs

Or from the dashboard: Settings > Context Injection Logs.

Each log entry shows:

• The user prompt that triggered the hook
• Which knowledge entries were searched
• Which entries were injected
• Time taken for the search

Search Architecture

lm-assist uses a hybrid search approach for high-quality results:

Search Flow

User Query
    |
    +----> Vector Search (semantic)
    |           |
    +----> BM25 Search (keyword)
    |           |
    v           v
    RRF Merge (fuse rankings)
         |
         v
    Content-match boost
         |
         v
    Recency weighting
         |
         v
    Top N results returned

Knowledge Quality

Knowledge quality improves over time through:

1. Feedback loop -- mark entries as outdated, wrong, or useful
2. Deduplication -- overlapping entries are consolidated
3. Recency -- newer knowledge ranks higher than stale entries
4. Project scoping -- knowledge is tagged with project paths for relevance

Configuration

Knowledge settings are available in the dashboard under Settings:

Cost Savings

The knowledge base pays for itself after just a few sessions by eliminating redundant codebase exploration.