go-palace

README

go-palace

Local-first memory palace for AI agents. Mine project files and conversations into a semantic store, search with vector similarity, and serve memories over MCP.

What it does

• Mine project files and conversation logs into structured memories with automatic room/wing classification
• Search semantically using sqlite-vec vector embeddings (offline, no API calls)
• Stack memories in 4 layers: identity, essential story, on-demand, deep search
• Serve 19 MCP tools for Claude Desktop, picoclaw, or any MCP-compatible client
• Graph knowledge as subject-predicate-object triples with timeline queries

Quick start

# Build (requires Go 1.26+, CGO for sqlite-vec)
make build

# Initialize a project — detects rooms from folder structure
bin/mempalace init ./my-project

# Mine project files into the palace
bin/mempalace mine ./my-project

# Search your memories
bin/mempalace search "authentication flow"

# Get a wake-up summary (L0 + L1 memory stack)
bin/mempalace wake-up

Architecture

cmd/mempalace/          CLI entry (cobra)
pkg/                    Public API — importable by consumers
  palace/               sqlite-vec vector store
  config/               JSON config + env vars
  embed/                Embedder interface (Hugot, Fake)
  searcher/             Semantic search
  layers/               L0-L3 memory stack
  kg/                   Knowledge graph (subject-predicate-object)
  sanitizer/            Input sanitization
internal/               Private implementation
  miner/                Project file mining (worker pool, gitignore)
  convominer/           Conversation mining
  normalize/            Chat format normalization (Claude/Codex/ChatGPT/Slack)
  extractor/            5-type heuristic memory extraction
  dialect/              AAAK dialect encoder/decoder
  graph/                Palace graph traversal
  entity/               Entity detection + registry
  room/                 Room detection from folders
  spellcheck/           Spell correction
  splitter/             Mega-file splitting
  hooks/                Harness hook handlers
  dedup/                Duplicate detection
  repair/               Palace repair utilities
  instructions/         Instruction handling
mcp/                    MCP JSON-RPC server over stdio
version/                Version constant

Library usage

go-palace is designed to be imported as a library. The pkg/ packages are the public API:

import (
    "go-palace/pkg/palace"
    "go-palace/pkg/embed"
    "go-palace/pkg/searcher"
    "go-palace/pkg/layers"
    "go-palace/pkg/kg"
)

// Open a palace with the default hugot embedder
emb, cleanup, err := embed.NewHugotEmbedder("/path/to/model")
defer cleanup()

p, err := palace.Open("/path/to/palace.db", emb)
defer p.Close()

// Add a memory
err = p.Upsert(palace.Drawer{
    ID:       "my-drawer-1",
    Document: "The auth module uses JWT tokens with 24h expiry.",
    Wing:     "technical",
    Room:     "backend",
})

// Search
results, err := searcher.Search(p, "authentication tokens", 5)

// Wake-up (L0 + L1 summary)
stack := layers.NewStack(p, configDir)
text, err := stack.WakeUp()

CLI commands

Command	Description
status	Print palace stats: drawers, wings, rooms
init [dir]	Detect rooms from folder structure, write mempalace.yaml
mine [path]	Mine project files or conversations into the palace
search [query]	Semantic search across all memories
wake-up	Generate L0+L1 wake-up text for agent context
split [dir]	Split mega-files (>300 LOC) into focused modules
hook run	Run a harness hook (reads JSON from stdin)
instructions [name]	Print built-in instruction sets
repair	Repair palace: fix orphaned drawers, rebuild indexes
dedup	Find and remove duplicate or near-duplicate drawers
compress	Compress drawers using AAAK dialect encoding
mcp	Start MCP server (JSON-RPC over stdio)

Mine modes

# Mine project files (default)
bin/mempalace mine ./my-project

# Mine conversations
bin/mempalace mine --mode convos ./conversations/

# Mine with general extraction (5-type heuristic)
bin/mempalace mine --mode convos --extract general ./conversations/

# Incremental — only re-mines changed files
bin/mempalace mine ./my-project

MCP server

go-palace exposes 19 tools via the Model Context Protocol. Configure in Claude Desktop:

{
  "mcpServers": {
    "mempalace": {
      "command": "/path/to/bin/mempalace",
      "args": ["mcp"]
    }
  }
}

Available MCP tools

Tool	Description
mempalace_status	Palace stats
mempalace_search	Semantic search
mempalace_add_drawer	Store a memory
mempalace_delete_drawer	Remove a memory
mempalace_check_duplicate	Check for duplicates before adding
mempalace_list_wings	List all wings
mempalace_list_rooms	List all rooms
mempalace_get_taxonomy	Get wing/room taxonomy
mempalace_get_aaak_spec	Get AAAK dialect specification
mempalace_traverse	Graph traversal from a drawer
mempalace_find_tunnels	Find cross-wing connections
mempalace_graph_stats	Palace graph statistics
mempalace_kg_query	Query knowledge graph
mempalace_kg_add	Add knowledge triple
mempalace_kg_invalidate	Invalidate a triple
mempalace_kg_timeline	Timeline of knowledge changes
mempalace_kg_stats	Knowledge graph statistics
mempalace_diary_write	Write a diary entry
mempalace_diary_read	Read diary entries

Configuration

Precedence

Environment variables > ~/.mempalace/config.json > compiled defaults.

Environment variables

Variable	Description
MEMPALACE_PALACE_PATH	Override palace database location
MEMPAL_PALACE_PATH	Alias for the above

Config file

~/.mempalace/config.json:

{
  "palace_path": "/custom/path/to/palace.db",
  "collection_name": "mempalace_drawers",
  "topic_wings": ["emotions", "consciousness", "memory", "technical", "identity", "family", "creative"],
  "hall_keywords": { "technical": ["code", "python", "api", "database"] }
}

Palace path

Default: ~/.mempalace/palace. Override with --palace flag, env var, or config file.

Embeddings

go-palace uses hugot for offline ONNX-based embeddings. No API calls, no network required after model download.

Model setup

On first run, hugot downloads the embedding model automatically. To pre-download or use a custom model path:

bin/mempalace mine --model /path/to/model ./my-project

For testing, a FakeEmbedder is available that produces deterministic vectors without loading a model.

Key types

Type	Package	Description
Palace	pkg/palace	sqlite-vec backed vector store for drawers
Drawer	pkg/palace	A memory unit: document + metadata (wing, room, source)
Embedder	pkg/embed	Interface for text-to-vector conversion
Stack	pkg/layers	4-layer memory stack (L0 identity through L3 deep search)
KG	pkg/kg	Knowledge graph with subject-predicate-object triples
Config	pkg/config	Runtime configuration
Searcher	pkg/searcher	Semantic search over palace drawers

Testing

make test           # Unit tests (~643 tests)
make test/race      # Tests with race detector
make test/suite     # Behavioral equivalence suite
make test/embed     # Hugot integration tests (requires model)
make test/semantic  # Semantic search quality tests (requires model)
make audit          # Full pipeline: tidy + vet + lint + test + build

Contributing

See CONTRIBUTING.md.

License

MIT

Credits

go-palace is a Go port of Python MemPalace — the original local-first memory palace for AI agents. The behavioral contracts, memory stack architecture, and AAAK dialect encoding are preserved from the Python implementation.