Goldie 🐕
A multi-agent memory MCP server written in Go that runs locally on your machine. Goldie stores typed, named memories in a shared SQLite vector index so multiple agents (Claude, Codex, etc.) can remember, recall, update_memory, and forget from one pool — replacing per-project MEMORY.md files with a backend they can all share.
Features
- Typed memories: Each memory has a type (
user,feedback,project,reference,opinion,idea,todo,reminder), a unique name, optional description, body, agent, and source - Shared pool: Scope is a SQLite file — point any number of agents at the same DB and they share memory
- Semantic recall: Filtered KNN over chunk embeddings; recall returns the parent memory plus the matched excerpt
- Graph recall: Memories get graph nodes; Goldie harvests stable concepts and facets from names and sources so broad recalls can return grouped context like
Acme->Acme Cloud - Multiple embedding backends: MiniLM (local via ONNX Runtime) or Ollama (any embedding model)
- File ingestion:
index_file/index_directoryimport files asreferencememories named by absolute path (checksum-gated upsert) - Async job queue: Long-running indexing operations and graph harvest/backfill jobs run in the background with progress tracking
Graph Recall
Goldie v4 adds a graph layer beside semantic search. Each memory has a memory node, and Goldie harvests concept nodes from descriptive memory names and file/source paths. Shared prefixes become facets through part_of edges, so memories like acme_cloud_passwordless_auth and acme_cloud_pricing can group under Acme Cloud, which itself belongs to Acme.
recall remains backward-compatible: it still returns the flat results list. When a query resolves to a concept, the response also includes concept_recall with the matched concept, facet groups, and directly attached memories. Clients that ignore concept_recall continue to work.
When a query contains a known concept or facet, recall can also include automatic_recall: an inspectable graph path with branch scores, convergence status, and a stop reason. Ambiguous queries fall back instead of forcing a precise answer.
Graph harvest runs asynchronously. remember creates the memory and memory node, then queues a debounced graph_harvest job so concept extraction does not hold the write lock. Existing databases are backfilled through the same job queue.
Optional graph hints can reinforce the harvested graph. about attaches a memory to concept labels, aliases adds alternate names to the memory node, and links adds explicit graph edges as {relation, target} objects. Hint-origin edges and aliases are preserved across harvest.
Graph inspection and cleanup tools are available for maintenance: list_nodes, get_node (aliased as map), and explain_recall show what the graph is doing; link_memory manually attaches a memory to a concept; merge_nodes merges duplicate concepts with collision-safe edge and alias movement.
Upgrading Existing Databases
Goldie v4 upgrades existing SQLite databases in place. On first launch it adds the graph tables, memory node references, and concept-node vector index. Existing memories keep working immediately through normal vector recall while graph backfill runs asynchronously.
Before first v4 launch, back up the database if it contains important memories:
cp ~/.local/share/goldie/index.db ~/.local/share/goldie/index.db.v3-backup
Backfill jobs use the same async queue as file indexing. Use list_jobs and job_status to inspect progress.
Requirements
MiniLM Backend (requires ONNX Runtime)
# macOS
brew install onnxruntime
# Ubuntu/Debian
sudo apt install libonnxruntime-dev
# Fedora/RHEL
sudo dnf install onnxruntime-devel
# Arch Linux
sudo pacman -S onnxruntime
Ollama Backend
If you want to use Ollama instead of MiniLM, you only need Ollama installed:
# macOS
brew install ollama
# Linux
curl -fsSL https://ollama.com/install.sh | sh
# Pull an embedding model
ollama pull nomic-embed-text
Skip the ONNX Runtime installation below if you only plan to use Ollama.
Installation
From Releases (recommended)
Download a pre-built binary from the releases page:
# Example for macOS Apple Silicon
curl -LO https://github.com/srfrog/goldie-mcp/releases/latest/download/goldie-mcp-darwin-arm64
chmod +x goldie-mcp-darwin-arm64
mv goldie-mcp-darwin-arm64 ~/bin/goldie-mcp
The release binaries are ad-hoc codesigned for macOS and include the MiniLM model, so no additional downloads are required.
Build from Source
Requires Go 1.22+, CGO enabled, and Git LFS (the model file is stored with LFS):
git lfs install # if not already configured
git clone https://github.com/srfrog/goldie-mcp
cd goldie-mcp
make build
Configuration
Command Line Flags
Environment Variables
Supported Ollama Embedding Models
For other models, set OLLAMA_EMBED_DIMENSIONS to the model's output dimensions.
Usage with Claude Code
With MiniLM (default)
claude mcp add -s user -e GOLDIE_DB_PATH=~/.local/share/goldie/index.db goldie /path/to/goldie-mcp
Note: ONNXRUNTIME_LIB_PATH is optional if the library is in a standard location.
With Ollama
claude mcp add -s user -e GOLDIE_DB_PATH=~/.local/share/goldie/index.db goldie /path/to/goldie-mcp -- -b ollama
Note: Make sure Ollama is running (ollama serve) before starting Claude Code.
Usage with Claude Desktop
Add to your Claude Desktop configuration (claude_desktop_config.json):
- macOS:
~/Library/Application Support/Claude/claude_desktop_config.json
With MiniLM (default)
{
"mcpServers": {
"goldie": {
"type": "stdio",
"command": "/path/to/goldie-mcp",
"env": {
"GOLDIE_DB_PATH": "/home/user/.local/share/goldie/index.db"
}
}
}
}
With Ollama
{
"mcpServers": {
"goldie": {
"type": "stdio",
"command": "/path/to/goldie-mcp",
"args": ["-b", "ollama"],
"env": {
"GOLDIE_DB_PATH": "/home/user/.local/share/goldie/index.db",
"OLLAMA_EMBED_MODEL": "nomic-embed-text"
}
}
}
}
Usage with OpenAI Codex
Add to your Codex configuration (~/.codex/config.toml):
With MiniLM (default)
[mcp_servers.goldie]
command = "/path/to/goldie-mcp"
[mcp_servers.goldie.env]
GOLDIE_DB_PATH = "/home/user/.local/share/goldie/index.db"
ONNXRUNTIME_LIB_PATH = "/path/to/libonnxruntime.so"
Note: ONNXRUNTIME_LIB_PATH is optional if the library is in a standard location. Homebrew will install it to /opt/homebrew/lib/libonnxruntime.dylib on macOS. In Linux, find it with ldconfig -p | grep onnxruntime.
With Ollama
[mcp_servers.goldie]
command = "/path/to/goldie-mcp"
args = ["-b", "ollama"]
[mcp_servers.goldie.env]
GOLDIE_DB_PATH = "/home/user/.local/share/goldie/index.db"
OLLAMA_EMBED_MODEL = "nomic-embed-text"
Memory Model
Goldie's index is a flat pool of memories. Each memory has these fields:
Sharing. "Scope" is just the SQLite file. Multiple agents pointed at the same GOLDIE_DB_PATH share the same pool of memories — there is no per-agent isolation. Use agent and source to filter on read/delete.
Naming and conflicts. Names must be unique. remember is strict — no upsert. If two agents try to create the same name, the second one gets an error and is expected to recall the existing memory and call update_memory (or pick a different name).
Update semantics. Name is immutable. update_memory accepts patches for type/description/body/source/agent; changes to description or body re-embed the chunks.
File ingestion. index_file / index_directory are the one exception to the no-upsert rule. They import files as memories of type=reference, with name = source = <absolute path>. Re-indexing the same path skips when the SHA-256 checksum matches and replaces the body when it doesn't.
Graph hints. Plain remember is enough for v4 graph recall: Goldie harvests concepts from memory names and file/source paths. Optional about, aliases, and links hints add user-provided graph structure when the name alone is not enough. Hinted graph data is preserved across future harvest runs.
Available Tools
remember
Create a new memory. Fails if name is already in use — recall it and use update_memory instead.
Parameters:
name(required): Unique identifier (e.g.,feedback_testing)type(required): One ofuser,feedback,project,reference,opinion,idea,todo,reminderbody(required): Full contentdescription(optional): One-line summaryagent(optional): Agent that created the memorysource(optional): Where the memory was generatedabout(optional): concept labels this memory is aboutaliases(optional): aliases for this memory nodelinks(optional): graph links as{relation, target}objects
recall
Hybrid recall over memories. Always returns the backward-compatible flat results list with the best matched excerpt for each memory. When the query resolves to graph concepts, the response may also include grouped concept_recall and inspectable automatic_recall metadata. Filter by type, agent, or source to narrow scope.
Parameters:
query(required): Topic or questionlimit(optional): Max results (default 5, max 20)include_body(optional): Include full memory bodies instead of metadata plus excerptstype,agent,source(optional): Filters
get_memory
Fetch a single memory's full record, including body, by id or name. Use this after recall when an excerpt is relevant and you need the full stored content.
Parameters:
id_or_name(required): Memory id or name
update_memory
Update an existing memory by id or name. Body/description changes re-embed.
Parameters:
id_or_name(required)type,description,body,source,agent(optional patches)about,aliases,links(optional graph hints)
forget
Delete memories. Requires at least one filter or a query — refuses to wipe everything. With a query, top matches within the (optional) filter are deleted.
Parameters:
name,type,agent,source(optional filters)query(optional): semantic matchlimit(optional): max matches when query is given (default 5)
list_memories
List memories matching the filter, newest first. Returns metadata only (no body).
Parameters:
type,agent,source(optional filters)limit(optional)
count_memories
Count memories matching the filter.
list_nodes
Inspect graph nodes for testing graph recall. Returns node labels plus alias and edge counts.
Parameters:
kind(optional): node kind such asconceptormemoryquery(optional): normalized label, alias, or memory-name substringlimit(optional): max nodes (default 50, max 200)
explain_recall
Trace hybrid recall for a query. Returns vector matches, grouped concept recall, automatic graph descent path, candidate scores, and stop reason.
Parameters:
query(required): topic or questionlimit(optional): max results per section (default 5, max 20)type,agent,source(optional filters)
get_node
Inspect one graph node with aliases plus incoming and outgoing edges. Aliased as map — same tool, shorter name for quick topic browsing.
Parameters:
id(optional): graph node idlabel(optional): exact node label when id is omitted; forkind=memory, memory name also workskind(optional): node kind when resolving by label (defaultconcept)
merge_nodes
Merge a duplicate concept into a canonical concept. Moves aliases and edges with collision-safe deduplication, then deletes the source node.
Parameters:
source(required): duplicate concept id, exact label, or aliastarget(required): canonical concept id, exact label, or alias
link_memory
Manually attach a memory to a concept. Creates the concept if the label is new.
Parameters:
memory(required): memory id or nameconcept(required): concept id, exact label, alias, or new concept labelrelation(optional):about,mentions, orcanonical_for(defaultabout)
index_file
Import a file as a reference memory. The memory's name is the absolute path; re-indexing updates in place when the checksum changes.
Parameters:
path(required)agent(optional): Agent that triggered the import
index_directory
Import every matching file in a directory as reference memories.
Parameters:
directory(required)pattern(optional, default*)recursive(optional, defaultfalse)agent(optional): Agent that triggered the import
job_status, list_jobs, clear_queue
Manage the async job queue. File indexing, graph harvest, and graph backfill jobs complete in the background; use job_status to check progress. list_jobs returns the newest jobs first, capped by limit (default 50, max 500), and reports the total matching count.
Skip Patterns
When indexing directories, Goldie automatically skips certain files and directories to avoid indexing irrelevant content.
Default Skip Patterns
If no .goldieskip file exists in the directory being indexed, Goldie uses these defaults:
Custom Skip Patterns
Create a .goldieskip file in the directory to define custom patterns. This replaces the defaults entirely. Same format as .gitignore, with the same pattern syntax.
# .goldieskip example
# Lines starting with # are comments
# Skip all dotfiles/dotdirs
.[!.]*
# Skip dependencies
node_modules/
vendor/
.venv/
# Skip build outputs
dist/
build/
target/
# Skip specific files
*.log
*.tmp
secrets.json
Pattern syntax:
*matches any sequence of characters?matches any single character[abc]matches any character in the set[!abc]matches any character NOT in the set- Patterns ending in
/match directories
Example Prompts
Example prompts you can use with Claude Code, Claude Desktop, or Codex:
remember
Remember as a feedback memory named "feedback_testing": don't mock the database in integration tests — we got burned last quarter by mock/prod divergence.
Remember this user fact named "user_role": senior Go engineer, ten years of experience, currently learning React.
Save an opinion named "ui_dark_mode": dark mode is easier on the eyes for long sessions.
recall
Recall what you know about database testing
Recall feedback memories about pull request size
What memories do I have about the API design?
Explain recall for "how does Acme handle auth"
List concept nodes matching Acme
update_memory
Update memory "feedback_testing": new body is "use a real Postgres in CI; the staging DB is reset nightly".
forget
Forget all opinion memories from agent "claude-opus-4-7"
Forget memories matching "old API design notes"
list_memories / count_memories
List my feedback memories
How many memories has agent "codex" written?
index_file / index_directory
Index the file ~/project/README.md
Index all *.md files in ~/docs recursively
Agent Configuration
Agents won't reach for goldie by default — Claude Code has its own /memory, and Codex has its own context handling. The repo ships two opinionated templates that nudge them toward the shared pool. Both are short and safe to drop in as-is.
Claude Code
Copy templates/CLAUDE.md to ~/.claude/CLAUDE.md (it's loaded into every Claude Code session):
cp templates/CLAUDE.md ~/.claude/CLAUDE.md
For project-scoped behavior instead, copy it to <project>/CLAUDE.md.
Codex (CLI / App)
Copy templates/AGENTS.md to ~/.codex/AGENTS.md (Codex loads AGENTS.override.md and AGENTS.md from $CODEX_HOME, default ~/.codex):
cp templates/AGENTS.md ~/.codex/AGENTS.md
As a belt-and-suspenders measure, append the following to ~/.codex/config.toml so the rule fires at session start even if Codex misses the AGENTS.md load:
developer_instructions = """
At session start, read and obey ~/.codex/AGENTS.md when it exists.
For persistent memory operations, prefer Goldie over local memory when the Goldie MCP server is connected.
"""
Customizing
Both templates are starting points. Edit them to:
- Restrict which memory
types the agent should create - Add project-specific naming conventions for
name(e.g.<area>_<topic>) - Override the default
agentvalue - Tighten or relax the "when NOT to use goldie" rules
Sharing One Database Across Machines
Goldie's "scope" is the SQLite file, so syncing the database file across machines (iCloud, Dropbox, Syncthing) gives you follow-me memory without running a server. The default journal mode (DELETE) is already safe to use under cloud sync — only one .db file exists, no WAL/SHM sidecars to get out of order. One caveat:
Don't write from two machines at once. Cloud sync is not a coordination layer. If two machines write while disconnected, the sync client picks a winner and the other side's writes are lost (or a conflict copy is created). Workflow: quit any goldie session before switching machines, let sync settle, start the new machine.
For real multi-writer multi-machine setups, run goldie on a server and connect via Tailscale, or use Litestream to stream WAL changes to S3/B2.
Opting into WAL for local-only DBs
If your DB lives on local disk (no cloud sync) and you want read-during-write concurrency under heavy multi-agent load, set GOLDIE_JOURNAL_MODE=WAL. The performance difference is negligible for typical memory-store workloads, but the option exists.
Replacing Per-Project MEMORY.md
Goldie is designed to replace the per-project MEMORY.md files that agents like Claude Code create on disk. Point every agent at the same GOLDIE_DB_PATH, instruct them to use remember / recall / update_memory / forget instead of file-based memory, and you get:
- One pool shared by Claude (any session, any project), Codex, and any other MCP-aware agent
- Semantic recall instead of file globbing
- Per-agent provenance via the
agentfield, queryable throughrecall/forgetfilters - Filtered cleanup (e.g. "forget all
feedbackmemories from agent X")
Indexing existing transcripts
You can also bulk-import old Claude Code conversation transcripts as reference memories so they participate in recall:
Index all *.md files in ~/.claude/projects recursively
Then ask:
Recall what I know about authentication bugs
What memories do I have about Docker?
Troubleshooting
macOS: Binary killed immediately (Signal 9)
When copying binaries on macOS, Gatekeeper may add quarantine attributes (com.apple.provenance) that cause the binary to be killed on launch. Use make install DEST=<path> which builds directly to the destination and codesigns the binary to avoid this issue.
Codex is not recalling
- MCP support in Codex is experimental and it needs a bit more coaxing to work properly. After indexing content, try to
recall <topic>and check that it uses thegoldie.recall()function; that indicates it's using the MCP backend. - Codex sometimes doesn't trust the content from recall and won't add it to the context. Try
recall <topic> and consolidateto push the update.
Architecture
goldie-mcp/
├── main.go # MCP server setup and tool handlers
├── internal/
│ ├── embedder/ # Embedding interface and backends
│ │ ├── minilm/ # MiniLM backend (ONNX Runtime)
│ │ └── ollama/ # Ollama backend (API client)
│ ├── goldie/ # Memory operations (Remember/Recall/Update/Forget)
│ │ ├── goldie.go # Core, file ingestion, chunking
│ │ └── memory.go # Type whitelist + memory CRUD
│ ├── store/ # SQLite memory + chunk + vec storage
│ │ ├── store.go # Connection, jobs
│ │ └── memory.go # Memory schema and queries
│ └── queue/ # Async job processing
├── go.mod
└── Makefile
Schema
Core SQLite tables make up the memory index:
memories— one row per memory:id, name UNIQUE, type, description, body, agent, source, checksum, created_at, updated_atmemory_chunks— body split into overlapping chunks for embedding granularity:id, memory_id, chunk_index, contentmemories_vec—vec0virtual table over chunk embeddings, joined back to memories on recallnodes,node_aliases,edges— graph layer for memory nodes, harvested concepts, aliases, and relationshipsnodes_vec—vec0virtual table over concept node embeddings for fuzzy concept resolution
Recall does KNN over chunks, dedupes to distinct memories, and then augments the flat results with graph groups and automatic graph-descent metadata when the query matches known concepts.
Embedding Backends
MiniLM Backend (-b minilm)
Uses all-MiniLM-L6-v2 via ONNX Runtime:
- 384-dimensional embeddings
- Optimized for semantic similarity
- Runs locally, model embedded in binary
Ollama Backend (-b ollama)
Uses Ollama's embedding API with your choice of model:
nomic-embed-text(768 dimensions) - Default, good balance of quality and speedmxbai-embed-large(1024 dimensions) - Higher quality embeddingsall-minilm(384 dimensions) - Same model as MiniLM backend- Any other Ollama embedding model (set
OLLAMA_EMBED_DIMENSIONS)
Note: Different embedding models produce different dimension vectors. Memories indexed with one backend/model cannot be recalled using another with different dimensions. Use separate databases or re-index when switching.
License
MIT
