muninn-remembers 0.1.6

Semantic memory layer for OpenCode AI coding sessions

Muninn

Muninn (Old Norse: "memory") is one of Odin's two ravens — the raven of memory.

Muninn is a semantic memory layer for OpenCode. It gives AI coding sessions persistent, searchable memory — so you never have to re-explain your project from scratch.

What it does

• Session resumption — load prior context at the start of every session
• Decision tracking — record architectural decisions with their rationale
• Pattern memory — save code conventions that apply across the project
• Semantic search — find relevant memories by meaning, not exact keywords
• Global memory — store cross-project knowledge (infra procedures, tool patterns, auth flows) that persists across all projects
• Symbol index — index code symbols (functions, classes, methods) and search them semantically so the AI can navigate large codebases without re-reading files

Memory is stored locally in ~/.config/opencode/muninn/chroma/ using ChromaDB (embedded). Each project gets its own isolated collection. A special __global__ collection holds cross-project knowledge.

---

Prerequisites

1. uv — https://docs.astral.sh/uv/
2. Ollama with mxbai-embed-large — can be local or remote:
   • Local: ollama pull mxbai-embed-large && ollama serve
   • Remote: any Ollama-compatible endpoint (e.g. Mimir) — set MUNINN_OLLAMA_URL and optionally MUNINN_OLLAMA_TOKEN

---

Installation

OpenCode

1. Install companion skills

uvx muninn-remembers install opencode

This copies memory-read, memory-write, and symbol-search to ~/.config/opencode/skills/.

2. Add the MCP server to ~/.opencode/opencode.json

"muninn": {
  "type": "local",
  "command": ["uvx", "muninn-remembers"],
  "environment": {
    "MUNINN_OLLAMA_URL": "http://localhost:11434",
    "MUNINN_DATA_DIR": "/Users/your-username/.config/opencode/muninn"
  },
  "enabled": true
}

3. Restart OpenCode to pick up the new MCP server.

---

Claude Code

1. Install companion slash commands

uvx muninn-remembers install claude

This copies memory-read.md, memory-write.md, and symbol-search.md to ~/.claude/commands/, making them available as /memory-read, /memory-write, and /symbol-search slash commands.

2. Add the MCP server

claude mcp add muninn \
  --env MUNINN_OLLAMA_URL=http://localhost:11434 \
  --env MUNINN_DATA_DIR=/Users/your-username/.config/opencode/muninn \
  -- uvx muninn-remembers

3. Verify by running /mcp in Claude Code — muninn should appear in the list.

---

Environment Variables

Variable	Default	Description
MUNINN_OLLAMA_URL	http://localhost:11434	Ollama base URL (local or remote)
MUNINN_OLLAMA_TOKEN	(unset)	Bearer token for authenticated Ollama endpoints (e.g. Mimir)
MUNINN_EMBED_MODEL	mxbai-embed-large:latest	Ollama embedding model
MUNINN_DATA_DIR	~/.config/opencode/muninn	ChromaDB storage path
MUNINN_TOP_K	5	Default number of search results
MUNINN_PROJECT	(auto-detected)	Override project name

Using a remote Ollama endpoint (e.g. Mimir)

OpenCode (opencode.json):

"muninn": {
  "type": "local",
  "command": ["uvx", "muninn-remembers"],
  "environment": {
    "MUNINN_OLLAMA_URL": "https://your-mimir-host/v1",
    "MUNINN_OLLAMA_TOKEN": "your-bearer-token",
    "MUNINN_DATA_DIR": "/Users/your-username/.config/opencode/muninn"
  },
  "enabled": true
}

Claude Code:

claude mcp add muninn \
  --env MUNINN_OLLAMA_URL=https://your-mimir-host/v1 \
  --env MUNINN_OLLAMA_TOKEN=your-bearer-token \
  --env MUNINN_DATA_DIR=/Users/your-username/.config/opencode/muninn \
  -- uvx muninn-remembers

When MUNINN_OLLAMA_TOKEN is set, every embedding request includes an Authorization: Bearer <token> header.

---

MCP Tools

Project-scoped tools

Tool	Description
memory_write	Write a memory entry (text, type, tags)
memory_search	Semantic search across project memories
memory_list	List recent memories (paginated)
memory_delete	Delete a memory entry by ID
memory_wipe_project	Delete ALL memories for a project (requires confirm=True)
memory_list_projects	List all projects that have stored memories

Global tools (cross-project)

Tool	Description
global_memory_write	Write a cross-project memory (infra procedures, tool patterns, auth flows, etc.)
global_memory_search	Semantic search across global memories
global_memory_list	List global memories (paginated)
global_memory_delete	Delete a global memory entry by ID
global_memory_wipe	Delete ALL global memories (requires confirm=True)

Symbol tools (per-project code index)

Tool	Description
symbol_index	Index one or more code symbols — name, kind, file, signature, docstring, callers
symbol_search	Semantic search across indexed symbols by natural-language query
symbol_delete_file	Remove all symbols belonging to a given file path
symbol_wipe	Delete the entire symbol index for a project (requires confirm=True)

Each symbol is stored in a dedicated ChromaDB collection named <project>__symbols, separate from the memory collection. Symbol entries are keyed by a deterministic ID (sha1(<project>:<file>:<name>:<kind>)[:16]), so re-indexing a symbol overwrites the previous entry rather than creating a duplicate.

---

Skills

Two companion skills guide the agent:

• memory-read — Load project context + global context at session start
• memory-write — Save decisions, summaries, next steps, and cross-project knowledge
• symbol-search — Guide when and how to index code symbols and search them by meaning

---

Verify it works

After installation, call the memory_list_projects tool (in OpenCode or Claude Code):

memory_list_projects()

This returns a (possibly empty) list if Muninn is connected. If you see an error, check that:

1. uv is on your PATH
2. Ollama is reachable: curl $MUNINN_OLLAMA_URL/api/tags (or curl http://localhost:11434/api/tags for local)
3. The MCP server is registered — run /mcp in Claude Code, or check opencode.json in OpenCode

---

Project detection

Muninn auto-detects the current project using this priority order:

1. MUNINN_PROJECT env var (explicit override)
2. git rev-parse --show-toplevel → basename
3. os.getcwd() → basename

---

Data location

All memory data lives at ~/.config/opencode/muninn/chroma/. To back up your memories, copy this directory. To start fresh for a project, use memory_wipe_project.

---

Publishing

Releases are published to PyPI automatically when a v* tag is pushed:

git tag v0.1.0
git push --tags

One-time PyPI setup (maintainer only):

1. Create an account at pypi.org
2. Go to Account → Publishing → Add a Trusted Publisher:
   • Owner: antic-eye
   • Repository: your GitHub repo name
   • Workflow: publish.yml
   • Environment: (leave blank)