Power your agent
การติดตั้งจากผู้เผยแพร่ ไม่ใช่เอนด์พอยต์ของ PlugMCP
URL แบบรีโมตหรือคำสั่งบนเครื่องของคุณ? การติดตั้งสองแบบต่างกันอย่างไร
รันบน PlugMCP
A new server you create. It is not this listing's official endpoint. Use the publisher install when you want their URL and their credentials.
เปิดจากผู้เผยแพร่
Deterministic local memory for LLM agents.
tagmem is a local-first memory system that stores original text, retrieves it with hybrid semantic and lexical ranking, and exposes that memory through a simple CLI and MCP interface.
Install · MCP · Benchmarks · · ·
Structured local memory storage and retrieval for LLM agents
Install tagmem with one command:
curl -fsSL https://raw.githubusercontent.com/codysnider/tagmem/main/scripts/install.sh | bash
For a detailed installation guide, see INSTALL.md.
tagmem is built for teams and individuals who want memory behavior they can inspect, reproduce, and trust.
Published images:
ghcr.io/codysnider/tagmem:latest-cpu
ghcr.io/codysnider/tagmem:latest-gpu
After installation, just use tagmem.
Local storage is created automatically on first use.
The installer chooses latest-gpu when the GPU image validates successfully on an NVIDIA host. Otherwise it installs wrappers for latest-cpu.
If OpenCode is detected during an interactive install and its config is patchable, the installer can offer to add the tagmem MCP entry.
Source build means compiling from the repository yourself. The simple installer path is Docker-only for now, and this audit pass live-validated the Linux/NVIDIA Docker path only.
Add an entry:
tagmem add --depth 0 --title "identity" --body "You are building a local memory system."
Search:
tagmem search "identity"
tagmem search --depth 2 "auth migration"
tagmem search --tag auth "token refresh"
tagmem search --explain "What database does staging use?"
tagmem uses a simple memory model:
This keeps the memory layer understandable while still supporting richer retrieval behavior.
For a practical rule set on when text should stay an entry, become a fact, or both, see FACT_RUBRIC.md. For the graph model and query semantics, see KNOWLEDGE_GRAPH.md.
Core commands:
tagmem inittagmem ingesttagmem splittagmem addtagmem listtagmem searchtagmem showtagmem statustagmem contexttagmem depthstagmem pathstagmem doctortagmem servetagmem repairtagmem mcptagmem benchtagmem init remains available as an optional bootstrap command if you want to precreate storage and print the resolved paths.
Examples:
tagmem ingest --mode files --depth 1 ~/projects/my_app
tagmem ingest --mode conversations --depth 2 ~/chats
tagmem ingest --mode conversations --extract general ~/chats
tagmem split ~/chats
tagmem status
tagmem context --depth 0
tagmem context --tag auth
tagmem show 1
Run the MCP server over stdio:
tagmem mcp
If a reachable local daemon socket already exists, tagmem mcp will use it automatically. To make daemon-backed MCP required instead of opportunistic, run the local daemon first and set TAGMEM_MCP_USE_DAEMON=1:
tagmem serve
TAGMEM_MCP_USE_DAEMON=1 tagmem mcp
Current MCP tools:
tagmem_statustagmem_pathstagmem_list_depthstagmem_list_tagstagmem_get_tag_maptagmem_list_entriestagmem_searchtagmem_fact_rubrictagmem_show_entrytagmem_check_duplicatetagmem_add_entrytagmem_delete_entrytagmem_kg_querytagmem_kg_addtagmem_kg_invalidatetagmem_kg_timelinetagmem_kg_statstagmem_graph_traversetagmem_find_bridgestagmem_graph_statstagmem_diary_writetagmem_diary_readtagmem_doctorCurrent benchmark snapshot:
Latest release-check confirmation for the default embedded model (bge-small-en-v1.5, ONNX/CUDA):
xychart-beta
title "LongMemEval Recall@5"
x-axis ["bge-base", "bge-small", "MemPalace"]
y-axis "Recall@5" 0.90 --> 1.00
bar [0.992, 0.990, 0.966]
tagmem (bge-small-en-v1.5): Recall@1 0.924, Recall@5 0.990, MRR 0.955tagmem (bge-base-en-v1.5): Recall@1 0.922, Recall@5 0.992, MRR 0.953Recall@5 0.966The latest guarded rerun for bge-small-en-v1.5 also recorded Recall@10 0.996, NDCG@10 0.951, and Time 25.1s.
Those guarded numbers track the benchmark CLI's component path. The repository also supports an interface path that exercises the real repository and search flow (TAGMEM_BENCH_PATH=interface or both); that path is reported separately because its latency and recall differ from the direct harness path. When a reachable local daemon socket is present, the interface path may reuse the daemon's hot corpus state instead of rebuilding a local corpus for each run.
The latest audited warm interface rerun for bge-small-en-v1.5 recorded Recall@1 0.874, Recall@5 0.980, MRR 0.921, and Time 360.2s; the raw artifact is published separately at benchmarks/raw/bge-small-en-v1.5/longmemeval-interface.json.
FalseMemBench is a standalone adversarial distractor benchmark focused on conflicting, stale, and near-miss memories.
These values are currently published from the checked-in raw artifacts under benchmarks/raw/adversarial/. The standalone FalseMemBench harness is maintained outside this repository and was not rerun in this audit pass.
xychart-beta
title "FalseMemBench Recall@1"
x-axis ["tagmem", "BM25", "MemPalace", "Contriever", "Stella"]
y-axis "Recall@1" 0.40 --> 0.90
bar [0.8674, 0.6946, 0.6632, 0.6527, 0.4258]
xychart-beta
title "FalseMemBench MRR"
x-axis ["tagmem", "BM25", "MemPalace", "Contriever", "Stella"]
y-axis "MRR" 0.60 --> 0.95
bar [0.9288, 0.8278, 0.8154, 0.8049, 0.6465]
tagmem: Recall@1 0.8674, Recall@5 0.9983, MRR 0.9288BM25: Recall@1 0.6946, Recall@5 0.9930, MRR 0.8278Recall@1 0.6632, Recall@5 0.9948, MRR 0.8154Contriever: Recall@1 0.6527, Recall@5 0.9843, MRR 0.8049Stella: Recall@1 0.4258, Recall@5 0.9791, MRR 0.6465For methodology, machine specs, and raw benchmark outputs, see:
Default embedded configuration:
export TAGMEM_EMBED_PROVIDER=embedded
export TAGMEM_EMBED_MODEL=bge-small-en-v1.5
export TAGMEM_EMBED_ACCEL=auto
export TAGMEM_EMBED_PROVIDER=openai
export TAGMEM_OPENAI_MODEL=nomic-embed-text
export TAGMEM_OPENAI_BASE_URL=http://localhost:11434/v1
export TAGMEM_OPENAI_API_KEY=
~/.local/share/tagmem/store.db~/.local/share/tagmem/store.json (not the live authority and may be absent between rebuilds)~/.local/share/tagmem/sources/~/.local/share/tagmem/vector/~/.local/share/tagmem/knowledge.json~/.local/share/tagmem/diaries/~/.local/share/tagmem/models/~/.config/tagmem/~/.cache/tagmem/Structured local memory storage and retrieval for LLM agents
เครื่องมือ
ยังไม่มีรายการเครื่องมือ เซิร์ฟเวอร์แบบรีโมตจะได้รายการหลังการตรวจสอบสถานะครั้งแรก
FAQ
คัดลอกคอนฟิก stdio สำหรับ Cursor, Claude Desktop, Claude Code หรือ VS Code ได้ในหน้านี้ คอนฟิกนี้ใช้การติดตั้งของผู้เผยแพร่ ไม่ใช่พร็อกซีของ PlugMCP
ระเบียนนี้ไม่ได้ระบุตัวแปรสภาพแวดล้อมที่จำเป็น
Tagmem ทำงานแบบโลคัลผ่าน stdio
ตรวจสอบล่าสุดเมื่อ 2026-10-28 Structured local memory storage and retrieval for LLM agents
