Skip to main content
Glama
hatrieukim

lifelog

by hatrieukim
OverviewSchemaSourceRelated ServersDiscussionsAdmin
Python
Local

Server Configuration

Describes the environment variables required to run the server.

NameRequiredDescriptionDefault
GRAPHNoSet to none to turn Graphiti off; episodes then stop at indexed.
DIARIZERNoSet to none to skip speaker diarization, or pyannote to use it.
HF_TOKENNoHugging Face read token for pyannote speaker diarization. Optional for the MCP server; without it set DIARIZER=none.
API_TOKENNoToken required before the HTTP server is reachable from anything other than this Mac, e.g. the iPhone app. Not needed for the stdio MCP server.
ASR_BACKENDNoSpeech-to-text backend: gpt-transcribe, gemini-3.5-transcribe, or chirp_3.
GEMINI_MODELNoGemini model id sent to Google; use gemini-3.5-transcribe-preview if Gemini returns 403 or 404.
OPENAI_API_KEYNoOpenAI API key, needed only when using gpt-transcribe as the speech-to-text backend.
CLAUDE_CODE_PROJECTSNoComma-separated Claude Code project folder names the worker may read. Empty means nothing is read.
GOOGLE_CLOUD_PROJECTYesGoogle Cloud project used for Gemini/Chirp speech-to-text, embeddings and the Graphiti graph. Required for the MCP server's search, summary, people and timeline tools.

Instructions

Guidance the server publishes about itself, which clients place ahead of the tool catalog so the model reads it before choosing anything.

This server publishes no instructions, or was last inspected before Glama recorded them.

Capabilities

Features and capabilities supported by this server

Protocol revision2025-11-25

CapabilityDetails
tools
{
  "listChanged": false
}
prompts
{
  "listChanged": false
}
resources
{
  "subscribe": false,
  "listChanged": false
}

Tools

Functions exposed to the LLM to take actions

NameDescription
searchA

Tìm trong bộ nhớ của người dùng: hội thoại đã ghi âm và các phiên làm việc với Claude Code. Trả về hai phần.

Nguyên văn: các đoạn gần nghĩa nhất với câu hỏi, có giờ, người nói, nguồn (ghi âm hay dự án nào), và độ gần
từ 0 tới 1. Độ gần dưới khoảng 0,5 thường là không liên quan. Với phiên Claude Code, đoạn ghi "tóm tắt cả phiên"
là bản tóm tắt do máy viết; các đoạn khác là lời thật của người dùng và của Claude trong phiên đó.
Hỏi việc đã làm ở một dự án thì nêu tên dự án trong `query` và dùng `since`, `until` để khoanh ngày.

Fact: những điều đã được chắt lọc từ hội thoại (ai, việc gì, ở đâu, quan hệ), kèm thời gian fact đó đúng.
Fact ghi "ĐÃ HẾT ĐÚNG" là điều từng đúng nhưng về sau đã đổi. Hội thoại không nhắc tên riêng nào thì
thường không có fact, chỉ có nguyên văn.

Mỗi kết quả ghi mã episode và số lượt nói để dẫn nguồn. Không thấy kết quả phù hợp thì nói là không
tìm thấy, đừng đoán.
summaryA

What the user did on one day or in one week, written from all sources: Claude Code sessions by project, recorded conversations and online meetings, the calendar, time spent in apps on the Mac, and pages read.

A day starts with an overview paragraph, then one section per source. A week has a summary by theme (main
work, results, people met, open items) and the overview of each day. Today and the current week are written
from the data so far. Every conversation and session carries its episode id, to look it up with `search`.
peopleB

Everything known about one person: how often and when their voice was recorded and who else was there, the latest conversations with them (with a short summary), calendar meetings they were invited to, chats with them on the Mac, their direct chats with the user on Zalo (how much, the groups they write in, the latest days with what each was about), what the memory has gathered about them, and the newest facts that involve them. Their Zalo name also finds them.

When several people match, the answer lists their names: call again with one of them.
timelineA

How a topic changed over time, oldest first, grouped by day: the facts about it in the order they became true, including those marked ĐÃ HẾT ĐÚNG (true once, replaced later), next to the passages where it was discussed. Use it for "how did we end up deciding X" or "what happened with Y over the last weeks".

Facts are picked by closeness in meaning, so a few may be only loosely related: ignore those. Each entry
carries its episode id and turn numbers for citing.

Prompts

Interactive templates invoked by user choice

NameDescription

No prompts

Resources

Contextual data attached and managed by the client

NameDescription

No resources

Deploy Server
Context All Your Agents Can ShareAgents share what they learn, hand off work, and hold it when you switch models. Start with 5 GB free.fulcradynamics.comAd

TDQS

A3.7/5.0

Scored across 4 tools

Disambiguation4/5

Each tool targets a distinct query intent: free-text retrieval (search), day/week digest (summary), entity lookup (people), and chronological topic evolution (timeline). search and timeline both surface facts and episodes, which could cause minor misselection, but their framing (query-driven vs. time-ordered) keeps them separable.

Naming Consistency4/5

All four names are single lowercase words with no mixed conventions (no camelCase/snake_case clashes), so the set reads predictably. They are not a strict verb_noun pattern and mix noun forms (people, timeline, summary) with a verb (search), a minor deviation.

Tool Count4/5

Four tools is on the lean side but each occupies a genuinely different retrieval axis for a personal memory domain. Nothing appears redundant, though a couple more query surfaces might be expected given the breadth of data sources described.

Completeness4/5

The read-only query surface covers search, aggregation, entity, and timeline views, which matches a passively-ingested lifelog where writes are automatic rather than tool-driven. Minor gaps exist (e.g., no direct fetch-by-episode-id, no explicit listing of sources/projects), but workable via search.

Maintenance

ActivityMaintained
ResponsivenessNo issues

Resources