AGENTS.md - Basic Memory Project Guide

February 2, 2026 · View on GitHub

Project Overview

Basic Memory is a local-first knowledge management system built on the Model Context Protocol (MCP). It enables bidirectional communication between LLMs (like Claude) and markdown files, creating a personal knowledge graph that can be traversed using links between documents.

CODEBASE DEVELOPMENT

Project information

See the README.md file for a project overview.

Build and Test Commands

  • Install: just install or pip install -e ".[dev]"
  • Run all tests (SQLite + Postgres): just test
  • Run all tests against SQLite: just test-sqlite
  • Run all tests against Postgres: just test-postgres (uses testcontainers)
  • Run unit tests (SQLite): just test-unit-sqlite
  • Run unit tests (Postgres): just test-unit-postgres
  • Run integration tests (SQLite): just test-int-sqlite
  • Run integration tests (Postgres): just test-int-postgres
  • Run impacted tests: just testmon (pytest-testmon; only tests affected by changed code)
  • Run MCP smoke test: just test-smoke
  • Fast local loop: just fast-check (default iteration flow)
  • Local consistency check: just doctor
  • Run all consolidated agent package checks: just package-check
  • Run Claude Code plugin checks: just package-check-claude-code
  • Run shared skills checks: just package-check-skills
  • Run Hermes plugin checks: just package-check-hermes
  • Run OpenClaw plugin checks: just package-check-openclaw
  • Run host-native agent harness checks: just agent-harness-check
  • Generate HTML coverage: just coverage
  • Single test: pytest tests/path/to/test_file.py::test_function_name
  • Run benchmarks: pytest test-int/test_sync_performance_benchmark.py -v -m "benchmark and not slow"
  • Lint: just lint or ruff check . --fix
  • Type check: just typecheck or uv run ty check src tests test-int
  • Type check (pyright): just typecheck-pyright or uv run pyright
  • Format: just format or uv run ruff format .
  • Run all code checks: just check (runs lint, format, typecheck, test)
  • Create db migration: just migration "Your migration message"
  • Run development MCP Inspector: just run-inspector

Note: Project requires Python 3.12+ (uses type parameter syntax and type aliases introduced in 3.12)

Postgres Testing: Uses testcontainers which automatically spins up a Postgres instance in Docker. No manual database setup required - just have Docker running.

Doctor Note: just doctor runs with a temporary HOME/config so it won't touch your local Basic Memory settings. It leaves temp dirs in /tmp (safe to ignore or remove).

Testmon Note: When no files have changed, just testmon may collect 0 tests. That's expected and means no impacted tests were detected.

Code/Test/Verify Loop (fast path)

  1. Code: make changes.
  2. Test: just fast-check (lint/format/typecheck + pytest-testmon impacted tests for changed code).
  3. Verify: just doctor (end-to-end file ↔ DB loop in a temp project).
  4. Package verify: just package-check when changes touch plugins/, skills/, integrations/, package metadata, or release wiring.
  5. Full gate (when needed): just test or just check for SQLite + Postgres.

Run just test-smoke when you specifically need the MCP smoke flow.

If testmon is “cold,” the first run may be long. Subsequent runs get much faster.

Consolidated Agent Package Checks

The monorepo ships several host-native packages alongside the Python core. Use the root justfile as the canonical entry point:

  • just package-check — validates every copied package and generated bundle path.
  • just package-check-claude-code — validates the root and plugin-local Claude marketplace manifests, the SessionStart/PreCompact hooks, the bundled output style, and the seed schemas, then runs claude plugin validate . --strict.
  • just package-check-skills — validates every top-level skills/memory-*/SKILL.md frontmatter block.
  • just package-check-hermes — validates integrations/hermes/plugin.yaml, the Hermes provider entrypoint, bundled skill, and runs the hermetic unit suite.
  • just package-check-openclaw — runs the OpenClaw package install, copies top-level skills into the generated bundle, typechecks, lints, builds dist/, runs Bun tests, and performs npm pack --dry-run.
  • just agent-harness-check — checks the host-specific harnesses without the shared markdown-only skills target.

Package-local justfiles live in plugins/claude-code/, skills/, integrations/hermes/, and integrations/openclaw/. Prefer the root targets for PR verification so command names stay stable as package internals evolve.

PR CI Gate

Before opening or updating a PR, run the checks that mirror the common required CI failures:

  • Run just typecheck in addition to targeted ruff and pytest commands when tests were added or changed.
  • Sign commits with git commit -s so DCO passes. If a PR branch already has unsigned commits, rewrite the branch with signed-off commits before asking for review.
  • Use a semantic PR title accepted by .github/workflows/pr-title.yml: type(scope): summary.
  • Use one of the allowed scopes: core, cli, api, mcp, sync, ui, ci, deps, installer, plugins, skills, integrations.

Test Structure

  • tests/ - Unit tests for individual components (mocked, fast)
  • test-int/ - Integration tests for real-world scenarios (no mocks, realistic)
  • Both directories are covered by unified coverage reporting
  • Benchmark tests in test-int/ are marked with @pytest.mark.benchmark
  • Slow tests are marked with @pytest.mark.slow
  • Smoke tests are marked with @pytest.mark.smoke

Code Style Guidelines

  • Line length: 100 characters max
  • Python 3.12+ with full type annotations (uses type parameters and type aliases)
  • Format with ruff (consistent styling)
  • Import order: standard lib, third-party, local imports
  • Naming: snake_case for functions/variables, PascalCase for classes
  • Prefer async patterns with SQLAlchemy 2.0
  • Use Pydantic v2 for data validation and schemas
  • CLI uses Typer for command structure
  • API uses FastAPI for endpoints
  • Follow the repository pattern for data access
  • Tools communicate to api routers via the httpx ASGI client (in process)

Code Change Guidelines

  • Full file read before edits: Before editing any file, read it in full first to ensure complete context; partial reads lead to corrupted edits
  • Minimize diffs: Prefer the smallest change that satisfies the request. Avoid unrelated refactors or style rewrites unless necessary for correctness
  • No speculative getattr: Never use getattr(obj, "attr", default) when unsure about attribute names. Check the class definition or source code first
  • Fail fast: Write code with fail-fast logic by default. Do not swallow exceptions with errors or warnings
  • No fallback logic: Do not add fallback logic unless explicitly told to and agreed with the user
  • No guessing: Do not say "The issue is..." before you actually know what the issue is. Investigate first.

Literate Programming Style

Code should tell a story. Comments must explain the "why" and narrative flow, not just the "what".

Section Headers: For files with multiple phases of logic, add section headers so the control flow reads like chapters:

# --- Authentication ---
# ... auth logic ...

# --- Data Validation ---
# ... validation logic ...

# --- Business Logic ---
# ... core logic ...

Decision Point Comments: For conditionals that materially change behavior (gates, fallbacks, retries, feature flags), add comments with:

  • Trigger: what condition causes this branch
  • Why: the rationale (cost, correctness, UX, determinism)
  • Outcome: what changes downstream
# Trigger: project has no active sync watcher
# Why: avoid duplicate file system watchers consuming resources
# Outcome: starts new watcher, registers in active_watchers dict
if project_id not in active_watchers:
    start_watcher(project_id)

Constraint Comments: If code exists because of a constraint (async requirements, rate limits, schema compatibility), explain the constraint near the code:

# SQLite requires WAL mode for concurrent read/write access
connection.execute("PRAGMA journal_mode=WAL")

What NOT to Comment: Avoid comments that restate obvious code:

# Bad - restates code
counter += 1  # increment counter

# Good - explains why
counter += 1  # track retries for backoff calculation

Codebase Architecture

See docs/ARCHITECTURE.md for detailed architecture documentation.

Directory Structure:

  • /alembic - Alembic db migrations
  • /api - FastAPI REST endpoints + container.py composition root
  • /cli - Typer CLI + container.py composition root
  • /deps - Feature-scoped FastAPI dependencies (config, db, projects, repositories, services, importers)
  • /importers - Import functionality for Claude, ChatGPT, and other sources
  • /markdown - Markdown parsing and processing
  • /mcp - MCP server + container.py composition root + clients/ typed API clients
  • /models - SQLAlchemy ORM models
  • /repository - Data access layer
  • /schemas - Pydantic models for validation
  • /services - Business logic layer
  • /sync - File synchronization services + coordinator.py for lifecycle management
  • /plugins/claude-code - Claude Code plugin marketplace package, hooks, skills, and agent harness
  • /skills - Canonical framework-agnostic Basic Memory SKILL.md source
  • /integrations/hermes - Hermes memory-provider plugin
  • /integrations/openclaw - OpenClaw npm/TypeScript plugin

Composition Roots: Each entrypoint (API, MCP, CLI) has a composition root that:

  • Reads ConfigManager (the only place that reads global config)
  • Resolves runtime mode via RuntimeMode enum (TEST > CLOUD > LOCAL)
  • Provides dependencies to downstream code explicitly

Typed API Clients (MCP): MCP tools use typed clients in mcp/clients/ to communicate with the API:

  • KnowledgeClient - Entity CRUD operations
  • SearchClient - Search operations
  • MemoryClient - Context building
  • DirectoryClient - Directory listing
  • ResourceClient - Resource reading
  • ProjectClient - Project management

Flow: MCP Tool → Typed Client → HTTP API → Router → Service → Repository

Development Notes

  • MCP tools are defined in src/basic_memory/mcp/tools/
  • MCP prompts are defined in src/basic_memory/mcp/prompts/
  • MCP tools should be atomic, composable operations
  • Use textwrap.dedent() for multi-line string formatting in prompts and tools
  • MCP Prompts are used to invoke tools and format content with instructions for an LLM
  • Schema changes require Alembic migrations
  • SQLite is used for indexing and full text search, files are source of truth
  • Testing uses pytest with asyncio support (strict mode)
  • Unit tests (tests/) use mocks when necessary; integration tests (test-int/) use real implementations
  • By default, tests run against SQLite (fast, no Docker needed)
  • Set BASIC_MEMORY_TEST_POSTGRES=1 to run against Postgres (uses testcontainers - Docker required)
  • Each test runs in a standalone environment with isolated database and tmp_path directory
  • CI runs SQLite and Postgres tests in parallel for faster feedback
  • Performance benchmarks are in test-int/test_sync_performance_benchmark.py
  • Use pytest markers: @pytest.mark.benchmark for benchmarks, @pytest.mark.slow for slow tests
  • Coverage must stay at 100%: Write tests for new code. Only use # pragma: no cover when tests would require excessive mocking (e.g., TYPE_CHECKING blocks, error handlers that need failure injection, runtime-mode-dependent code paths)

Async Client Pattern (Important!)

MCP tools use get_project_client() for per-project routing:

from basic_memory.mcp.project_context import get_project_client

@mcp.tool()
async def my_tool(project: str | None = None, context: Context | None = None):
    async with get_project_client(project, context) as (client, active_project):
        # client is routed based on project's mode (local ASGI or cloud HTTP)
        response = await call_get(client, "/path")
        return response

CLI commands and non-project-scoped code use get_client() directly:

from basic_memory.mcp.async_client import get_client

async def my_cli_command():
    async with get_client() as client:
        response = await call_get(client, "/path")
        return response

# Per-project routing (when project name is known):
async with get_client(project_name="research") as client:
    ...

Do NOT use:

  • ❌ from basic_memory.mcp.async_client import client (deprecated module-level client)
  • ❌ Manual auth header management
  • ❌ inject_auth_header() (deleted)
  • ❌ Separate get_client() + get_active_project() in MCP tools (use get_project_client() instead)

Key principles:

  • Auth happens at client creation, not per-request
  • Proper resource management via context managers
  • Per-project routing: each project can be LOCAL or CLOUD independently
  • Cloud projects use API key (cloud_api_key in config) as Bearer token
  • Routing priority: factory injection > force-local > per-project cloud > global cloud > local ASGI
  • Factory pattern enables dependency injection for cloud consolidation

For cloud app integration:

from basic_memory.mcp import async_client

# Set custom factory before importing tools
async_client.set_client_factory(your_custom_factory)

See SPEC-16 for full context manager refactor details.

Release Process

Releases are driven by just release / just beta — never by a bare git tag. The recipes bump version metadata, run pre-flight checks, commit, tag, and push. GitHub Actions then publishes to PyPI and updates the Homebrew formula.

Stable release:

just release v0.21.3

The recipe runs just lint + just typecheck, then updates every release manifest through scripts/update_versions.py: src/basic_memory/__init__.py, server.json, the root Claude marketplace, the Claude Code plugin manifest and local marketplace, the Hermes plugin.yaml, and the OpenClaw package.json. It commits as chore: update version to X.Y.Z for vX.Y.Z release, creates the vX.Y.Z tag, and pushes both the commit and the tag to origin/main. After the tag lands, the Release workflow builds the Python package, publishes to PyPI, creates the GitHub release with auto-generated notes, publishes the OpenClaw npm package, and updates the Homebrew formula. The recipe finishes by printing the post-release tasks the workflow doesn't cover.

Beta release: just beta v0.21.3b1 — same flow with a beta-suffixed tag. PyPI consumers install with pip install basic-memory --pre.

Release dry run: just release-dry-run v0.21.4 previews the consolidated version update without writing files.

Development builds: every commit to main publishes a 0.21.3.dev26+468a22f-style version to PyPI automatically via .github/workflows/dev-release.yml. No human action.

Do not tag releases by hand. A bare git tag vX.Y.Z skips the in-code version bump. Package metadata is still correct (uv-dynamic-versioning derives it from the git tag) but basic-memory --version reports the previous release, which is what happened with v0.21.2 → v0.21.3.

Post-release tasks the recipe surfaces but doesn't run:

  • docs.basicmemory.com — add notes to src/pages/latest-releases.mdx
  • basicmachines.co — bump version in src/components/sections/hero.tsx
  • MCP Registry — mcp-publisher publish from the repo root

See .claude/commands/release/release.md (and beta.md, release-check.md, changelog.md alongside it) for the full release + post-release runbook, including the slash commands.

BASIC MEMORY PRODUCT USAGE

Knowledge Structure

  • Entity: Any concept, document, or idea represented as a markdown file
  • Observation: A categorized fact about an entity (- [category] content)
  • Relation: A directional link between entities (- relation_type [[Target]])
  • Frontmatter: YAML metadata at the top of markdown files
  • Knowledge representation follows precise markdown format:
    • Observations with [category] prefixes
    • Relations with WikiLinks [[Entity]]
    • Frontmatter with metadata

Basic Memory Commands

Local Commands:

  • Check sync status: basic-memory status
  • Doctor check (file <-> DB loop): basic-memory doctor
  • Import from Claude: basic-memory import claude conversations
  • Import from ChatGPT: basic-memory import chatgpt
  • Import from Memory JSON: basic-memory import memory-json
  • Tool access: basic-memory tool (provides CLI access to MCP tools)
    • Continue: basic-memory tool continue-conversation --topic="search"

Project Management:

  • List projects: basic-memory project list
  • Add project: basic-memory project add "name" ~/path
  • Project info: basic-memory project info
  • Set cloud mode: basic-memory project set-cloud "name"
  • Set local mode: basic-memory project set-local "name"
  • One-way sync (local -> cloud): basic-memory project sync
  • Bidirectional sync: basic-memory project bisync
  • Integrity check: basic-memory project check

Cloud Commands (requires subscription):

  • Authenticate (global): basic-memory cloud login
  • Logout (global): basic-memory cloud logout
  • Check cloud status: basic-memory cloud status
  • Setup cloud sync: basic-memory cloud setup
  • Save API key: basic-memory cloud set-key bmc_...
  • Create API key: basic-memory cloud create-key "name"
  • Manage snapshots: basic-memory cloud snapshot [create|list|delete|show|browse]
  • Restore from snapshot: basic-memory cloud restore <path> --snapshot <id>

MCP Capabilities

  • Basic Memory exposes these MCP tools to LLMs:

    Content Management:

    • write_note(title, content, directory, tags) - Create/update markdown notes with semantic observations and relations
    • read_note(identifier, page, page_size) - Read notes by title, permalink, or memory:// URL with knowledge graph awareness
    • read_content(path) - Read raw file content (text, images, binaries) without knowledge graph processing
    • view_note(identifier, page, page_size) - View notes as formatted artifacts for better readability
    • edit_note(identifier, operation, content) - Edit notes incrementally (append, prepend, find/replace, replace_section)
    • move_note(identifier, destination_path, is_directory) - Move notes or directories to new locations, updating database and maintaining links
    • delete_note(identifier, is_directory) - Delete notes or directories from the knowledge base

    Knowledge Graph Navigation:

    • build_context(url, depth, timeframe) - Navigate the knowledge graph via memory:// URLs for conversation continuity
    • recent_activity(type, depth, timeframe) - Get recently updated information with specified timeframe (e.g., "1d", "1 week")
    • list_directory(dir_name, depth, file_name_glob) - Browse directory contents with filtering and depth control

    Search & Discovery:

    • search_notes(query, page, page_size, search_type, types, entity_types, after_date) - Full-text search across all content with advanced filtering options

    Project Management:

    • list_memory_projects() - List all available projects with their status
    • create_memory_project(project_name, project_path, set_default) - Create new Basic Memory projects
    • delete_project(project_name) - Delete a project from configuration

    Visualization:

    • canvas(nodes, edges, title, directory) - Generate Obsidian canvas files for knowledge graph visualization

    ChatGPT-Compatible Tools:

    • search(query) - Search across knowledge base (OpenAI actions compatible)
    • fetch(id) - Fetch full content of a search result document

  • MCP Prompts for better AI interaction:

    • ai_assistant_guide() - Guidance on effectively using Basic Memory tools for AI assistants
    • continue_conversation(topic, timeframe) - Continue previous conversations with relevant historical context
    • search(query, after_date) - Search with detailed, formatted results for better context understanding
    • recent_activity(timeframe) - View recently changed items with formatted output

Cloud Features (v0.15.0+)

Basic Memory now supports cloud synchronization and storage (requires active subscription):

Authentication:

  • JWT-based authentication with subscription validation
  • Secure session management with token refresh
  • Support for multiple cloud projects

Bidirectional Sync:

  • rclone bisync integration for two-way synchronization
  • Conflict resolution and integrity verification
  • Real-time sync with change detection
  • Mount/unmount cloud storage for direct file access

Cloud Project Management:

  • Create and manage projects in the cloud
  • Toggle between local and cloud modes
  • Per-project sync configuration
  • Subscription-based access control

Security & Performance:

  • Removed .env file loading for improved security
  • .gitignore integration (respects gitignored files)
  • WAL mode for SQLite performance
  • Background relation resolution (non-blocking startup)
  • API performance optimizations (SPEC-11)

Per-Project Cloud Routing:

Individual projects can be routed through the cloud while others stay local, using an API key:

# Save API key and set project to cloud mode
basic-memory cloud set-key bmc_abc123...
basic-memory project set-cloud research    # route through cloud
basic-memory project set-local research    # revert to local

MCP tools use get_project_client() which automatically routes based on the project's mode. Cloud projects use the cloud_api_key from config as Bearer token.

CLI Routing Flags (Global Cloud Mode):

When global cloud mode is enabled, CLI commands route to the cloud API by default. Use --local and --cloud flags to override:

# Force local routing (ignore cloud mode)
basic-memory status --local
basic-memory project list --local

# Force cloud routing (when cloud mode is disabled)
basic-memory status --cloud
basic-memory project info my-project --cloud

Key behaviors:

  • The local MCP server (basic-memory mcp) automatically uses local routing
  • This allows simultaneous use of local Claude Desktop and cloud-based clients
  • Some commands (like project default, project sync-config, project move) require --local in cloud mode since they modify local configuration
  • Environment variable BASIC_MEMORY_FORCE_LOCAL=true forces local routing globally
  • Per-project cloud routing via API key works independently of global cloud mode

AI-Human Collaborative Development

Basic Memory emerged from and enables a new kind of development process that combines human and AI capabilities. Instead of using AI just for code generation, we've developed a true collaborative workflow:

  1. AI (LLM) writes initial implementation based on specifications and context
  2. Human reviews, runs tests, and commits code with any necessary adjustments
  3. Knowledge persists across conversations using Basic Memory's knowledge graph
  4. Development continues seamlessly across different AI sessions with consistent context
  5. Results improve through iterative collaboration and shared understanding

This approach has allowed us to tackle more complex challenges and build a more robust system than either humans or AI could achieve independently.

Problem-Solving Guidance:

  • If a solution isn't working after reasonable effort, suggest alternative approaches
  • Don't persist with a problematic library or pattern when better alternatives exist
  • Example: When py-pglite caused cascading test failures, switching to testcontainers-postgres was the right call

GitHub Integration

Basic Memory has taken AI-Human collaboration to the next level by integrating Claude directly into the development workflow through GitHub:

GitHub MCP Tools

Using the GitHub Model Context Protocol server, Claude can now:

  • Repository Management:

    • View repository files and structure
    • Read file contents
    • Create new branches
    • Create and update files

  • Issue Management:

    • Create new issues
    • Comment on existing issues
    • Close and update issues
    • Search across issues

  • Pull Request Workflow:

    • Create pull requests
    • Review code changes
    • Add comments to PRs

This integration enables Claude to participate as a full team member in the development process, not just as a code generation tool. Claude's GitHub account (bm-claudeai) is a member of the Basic Machines organization with direct contributor access to the codebase.

Collaborative Development Process

With GitHub integration, the development workflow includes:

  1. Direct code review - Claude can analyze PRs and provide detailed feedback
  2. Contribution tracking - All of Claude's contributions are properly attributed in the Git history
  3. Branch management - Claude can create feature branches for implementations
  4. Documentation maintenance - Claude can keep documentation updated as the code evolves
  5. Code Commits: ALWAYS sign off commits with git commit -s
  6. Pull Request Titles: PR titles must follow the semantic format enforced by .github/workflows/pr-title.yml: type(scope): summary
    • Allowed types: feat, fix, chore, docs, style, refactor, perf, test, build, ci
    • Allowed scopes: core, cli, api, mcp, sync, ui, ci, deps, installer, plugins, skills, integrations
    • Example: fix(cli): propagate cloud workspace routing

This level of integration represents a new paradigm in AI-human collaboration, where the AI assistant becomes a full-fledged team member rather than just a tool for generating code snippets.