doobidoo/mcp-memory-service v10.29.0

v10.29.0 - LLM-based Harvest Classification via Groq

on GitHub

What's New in v10.29.0

This is a MINOR release adding an optional LLM-based classification layer to the session harvest tool.

Added

[#628] LLM-based classification layer for memory_harvest (Phase 2)

The memory_harvest MCP tool now accepts an optional use_llm boolean parameter. When enabled, extracted memories are routed through a new _GroqClassifierBridge that calls the Groq API for higher-precision category labels — especially valuable for nuanced or multi-topic memories where regex patterns are insufficient.

Key details:

• New harvest/classifier.py module with _GroqClassifierBridge: async-compatible, includes rate-limit handling, structured output parsing, and classify_batch() for efficient bulk classification
• Graceful fallback to rule-based classifier when use_llm=false (the default) or when the Groq API is unavailable (429/503)
• 14 new tests: classifier bridge mock responses, fallback paths, and end-to-end harvest flows
• Closes #618 — completes the two-phase harvest roadmap (Phase 1: rule-based extraction in #614; Phase 2: LLM re-classification)

Usage

# Basic harvest (unchanged, rule-based classifier)
memory_harvest(session_file="~/.claude/projects/my-project/session.jsonl")

# Phase 2: LLM-enhanced classification (requires GROQ_API_KEY)
memory_harvest(
    session_file="~/.claude/projects/my-project/session.jsonl",
    use_llm=True
)

Upgrading

This release is fully backward compatible. The use_llm parameter defaults to false, so existing integrations are unaffected.

pip install --upgrade mcp-memory-service==10.29.0
# or
uv add mcp-memory-service==10.29.0

Full Changelog: https://github.com/doobidoo/mcp-memory-service/blob/main/CHANGELOG.md#10290---2026-03-29