Smriti MCP

A portable memory server for AI agents, built for the Model Context Protocol (MCP).

Smriti stores durable memories as plain markdown files with YAML frontmatter. This keeps your data readable, git-friendly, and easy to inspect outside any single agent runtime.

Features

• Framework agnostic: Works with any MCP-compatible agent (Claude, OpenAI, local models, etc.)
• Durable & portable: All memories stored as plain markdown files—no database required
• Git-friendly: Version control your memories alongside your code
• Search & filter: Full-text search, filtering by tags, categories, and status
• Relationship tracking: Use [[wikilinks]] to connect related memories
• Memory index: Auto-generate markdown indexes of your entire memory store
• Archive & organize: Hierarchical organization with categories and status tracking

Installation

Smriti MCP is published on PyPI as smriti-mcp.

With pip

Install into your current Python environment:

pip install smriti-mcp

Then verify the CLI is available:

smriti-mcp --help

With uv

Install Smriti as a persistent command-line tool:

uv tool install smriti-mcp

Or run it directly without a separate install:

uvx smriti-mcp --help

From source for development

git clone https://github.com/deepak-bhardwaj-ps/smriti-mcp.git
cd smriti-mcp
pip install -e .

Quick Start

1. Run the server locally

smriti-mcp server --memory-root ~/.smriti/memory

By default, Smriti uses ~/.smriti/memory. You can override it with:

export SMRITI_MEMORY_ROOT="$HOME/.smriti/memory"
smriti-mcp server

If you prefer uvx, run the server with:

uvx smriti-mcp server --memory-root ~/.smriti/memory

2. Configure in your MCP client

Claude Desktop with pip or uv tool install (~/.config/claude_desktop_config.json):

{
  "mcpServers": {
    "smriti": {
      "type": "stdio",
      "command": "smriti-mcp",
      "args": ["server", "--memory-root", "~/.smriti/memory"]
    }
  }
}

Claude Desktop with uvx:

{
  "mcpServers": {
    "smriti": {
      "type": "stdio",
      "command": "uvx",
      "args": ["smriti-mcp", "server", "--memory-root", "~/.smriti/memory"]
    }
  }
}

Then restart Claude Desktop and Smriti will be available as a tool.

Available Tools

Core Operations

Tool	Description
create_memory	Create a new durable markdown memory with metadata
get_memory	Retrieve a memory by ID and return its full content
append_memory	Add content to the end of an existing memory
update_memory	Patch metadata or replace memory content
delete_memory	Permanently remove a memory
remember	Agent-friendly write API that records a trace and can create, append, or update
consolidate_memory	Create, append, or update a memory from reviewed trace content

Search & Browse

Tool	Description
search_memory	Full-text search across title, tags, categories, and body. Returns ranked results
list_memories	Browse memory metadata without loading full content. Filter by status, category, tags

Organization

Tool	Description
archive_memory	Mark a memory as archived (soft delete)
build_memory_index	Generate a markdown index of all memories for easy browsing
rebuild_memory	Fix frontmatter, apply/normalize wikilinks from titles and aliases, and rebuild indexes
load_memory_index	Load the generated index as markdown

Memory Format

Each memory is stored as a markdown file with YAML frontmatter:

---
id: project/Example Architecture Decision
title: Example Architecture Decision
category: project
tags:
  - architecture
  - decision
status: active
short_description: Decided to use async/await pattern
created_at: "2026-06-05T10:30:00+10:00"
updated_at: "2026-06-05T10:30:00+10:00"
---

## Background

We needed to handle concurrent requests efficiently.

## Decision

Use async/await with asyncio for I/O-bound operations.

## Consequences

- Improved throughput for concurrent operations
- Need to manage event loop carefully in multi-threaded contexts

See also: [[Async Migration]], [[Performance Metrics]]

Metadata Fields

• id: Unique identifier (auto-generated from category + title, or custom)
• title: Human-readable title
• category: Organizational category (becomes directory in file structure)
• tags: Array of searchable tags
• status: active, archived, or custom status
• short_description: Brief summary (used in indexes)
• created_at: ISO 8601 timestamp
• updated_at: ISO 8601 timestamp
• memory_type, confidence, salience, scope, source_agent: Optional agent memory metadata used for filtering and recall

File Structure

~/.smriti/memory/
├── project/
│   ├── Example Architecture Decision.md
│   ├── Async Migration.md
│   └── Performance Metrics.md
├── research/
│   └── LLM Benchmarks.md
├── decisions/
│   └── Use Postgres.md
└── index.md

Smriti keeps default filenames aligned with memory titles so Obsidian-style wikilinks like [[API Rate Limiting Strategy]] resolve to API Rate Limiting Strategy.md.

When you run rebuild_memory, Smriti can automatically add missing wikilinks and normalize alias links. It matches longer titles and aliases first and only links whole phrases, so Durable Memory is preferred over durable, and able is not linked inside durable.

Usage Examples

Create a memory

from smriti_mcp.store import MemoryStore

store = MemoryStore("~/.smriti/memory")

result = store.create_memory(
    {
        "title": "API Rate Limiting Strategy",
        "category": "decisions",
        "tags": ["api", "performance"],
        "short_description": "Decided on sliding window rate limiting",
    },
    content="We chose sliding window over token bucket because...",
)

# Returns: {"id": "decisions/API Rate Limiting Strategy", ...}

Remember with precise metadata

result = store.remember(
    content="User prefers markdown-first durable memory with no mandatory vector database.",
    id="preferences/Markdown First Memory",
    meta={
        "title": "Markdown First Memory",
        "category": "preferences",
        "memory_type": "preference",
        "short_description": "Preference for markdown-first durable memory.",
        "source_agent": "codex",
        "confidence": "high",
    },
    mode="create",
)

remember treats supplied meta as authoritative. If short_description is omitted, Smriti leaves it omitted instead of deriving a partial summary from the body. In auto mode, Smriti only appends to an existing memory when there is a strong deterministic match such as the same title; use id with append or update mode for explicit writes.

Search memories

results = store.search_memory(
    query="rate limiting",
    include_content=False,  # Just metadata
)

for result in results:
    print(f"{result['id']}: {result['title']}")

List memories with filters

active_decisions = store.list_memories(
    status="active",
    category="decisions",
)

for memory in active_decisions:
    print(f"{memory['title']} ({memory['status']})")

Rebuild and repair memories

result = store.rebuild_memory(
    fix_frontmatter=True,
    apply_wikilinks=True,
    group_by_category=True,
)

print(result["wikilinks"]["links_added"])

Running Tests

# Install test dependencies
pip install -e ".[dev]"

# Run all tests
pytest tests/ -v

# Run integration tests only
pytest tests/test_smriti_mcp_integration.py -v

All tests pass, including full MCP stdio round-trip integration tests.

Architecture

• MemoryStore: Core storage engine with markdown file I/O
• Server: MCP server exposing tools to agents
• CLI: Command-line interface for running the stdio server
• Frontmatter: YAML metadata parsing and generation

The package has zero external database dependencies and works with Python 3.10+.

Roadmap

• Web UI for browsing memories
• Multi-user support with authentication
• Memory graph visualization
• Sync to cloud storage (S3, GCS)
• Memory embeddings for semantic search

Contributing

Contributions are welcome! Please:

1. Fork the repository
2. Create a feature branch (git checkout -b feature/my-feature)
3. Add tests for new functionality
4. Ensure all tests pass (pytest tests/ -v)
5. Submit a pull request

License

MIT License - see LICENSE file for details.

Author

Created by Deepak Bhardwaj.

See Also

• Model Context Protocol (MCP)
• Claude MCP Documentation
• Smriti concept