greeum 3.1.1rc8

Context-Dependent Neural Memory System for LLMs with GraphIndex beam search, STM/LTM consolidation, and semantic tagging

Greeum

Context-aware memory for AI workflows. Keep long-running conversations and project notes persistent across tools.

English · 한국어

---

1. Installation & Setup

First run checklist

1. Install the package (pipx or pip)
2. Run greeum setup --start-worker to create the data directory and launch the worker
3. Connect your MCP client (Codex, ClaudeCode, Cursor, …)

👉 Need the ultra-short version? See docs/QUICKSTART.md for “설치 → 셋업 → 연동” 한 페이지 요약.

# Recommended (isolated) install
pipx install --pip-args "--pre" greeum

# or standard pip
pip install --upgrade "greeum"

# initialise data directory, choose where memories live
greeum setup --start-worker

Optional: enable semantic embeddings

pip install sentence-transformers          # once per machine
greeum mcp warmup                          # downloads the default model

• MCP/CLI run with hash fallback by default for fast startup.
• Add --semantic (or unset GREEUM_DISABLE_ST) when you want the SentenceTransformer-enabled search:

  greeum mcp serve --semantic -t stdio
  

---

2. MCP Integration

Codex (STDIO)

1. Ensure greeum setup has been run at least once.
2. ~/.codex/config.toml

   [mcp_servers.greeum]
   command = "greeum"
   args    = ["mcp", "serve", "-t", "stdio"]
   env     = { "GREEUM_QUIET" = "true", "PYTORCH_ENABLE_MPS_FALLBACK" = "1" }
   

3. Optional semantic mode:

   args = ["mcp", "serve", "-t", "stdio", "--semantic"]
   

   First run may take longer while the model loads. Warm-up before enabling for smoother startup.

ClaudeCode / Cursor (native MCP)

greeum mcp serve

• Add the command above to the client’s MCP configuration.
• Semantic mode: greeum mcp serve --semantic

HTTP / URL-based MCP (e.g. ChatGPT)

greeum mcp serve -t http --host 0.0.0.0 --port 8800

Then register http://127.0.0.1:8800/mcp as the endpoint.

---

3. LLM Prompting Guidelines

• Always close sessions with a summary: “Call add_memory summarising decisions before ending the shift.”
• Retrieve before writing: run search_memory with the task keywords before starting work.
• Use anchor slots (A/B/C) for hot contexts:

  {
    "name": "search_memory",
    "arguments": { "query": "login flow", "limit": 5, "slot": "A" }
  }
  

• Encourage agents to log important facts with importance ≥ 0.6 so team hand‑offs stay seamless.

---

4. CLI Essentials

# Add context
greeum memory add "Legal copy updated for release"

# Search (global fallback enabled by default)
greeum memory search "release notes" --count 5

# Anchor-based search (slot-aware)
greeum memory search "translations" --slot B --radius 2

# Rebuild branch indices (FAISS + keyword or keyword-only)
greeum memory reindex             # uses FAISS if available
greeum memory reindex --disable-faiss

# Reuse the long-running worker (avoids cold-start on each CLI call)
greeum worker serve --host 127.0.0.1 --port 8800   # terminal 1
export GREEUM_MCP_HTTP="http://127.0.0.1:8800/mcp" # terminal 2
greeum memory add "Sprint hand-off" --use-worker
greeum memory search "hand-off" --use-worker

Other useful commands:

• greeum anchors status / set A <block> / pin A
• greeum workflow search "<topic>" for scripted MCP calls
• greeum mcp warmup to cache the embedding model before enabling semantic mode

---

5. Documentation

• Getting Started
• MCP Integration Details
• Automation Workflow Guide
• API Reference

---

6. License

MIT License — see LICENSE.

---

Greeum · Persistent memory for AI—built and maintained by the community.