Local markdown memory & cross-agent context engine for AI coding assistants.
One persistent identity, shared across Claude Code, Cursor, Antigravity, and Zed.
The name agents-memory comes directly from its universal storage standard: .agents/memory.
- Global:
~/.agents/memory/stores your persistent identity, stack defaults, durable concepts, and project index. - Repository:
<repo>/.agents/memory/stores repo-specific facts, architecture ADRs, and staging inboxes.
While AI vendors fragment their configuration across proprietary stores, .agents/ provides a single, open, vendor-neutral filesystem hub for all agent configurations and shared intelligence.
agents-memory delivers this with a pure Markdown-first architecture:
- Local & Offline: Your identity and repo memory live in plain files (
~/.agents/memory/and<repo>/.agents/memory/). - Human-Readable & Git-Friendly: Edit with any text editor, diff with git, commit when you want.
- Universal MCP Server: Exposes memory tools (
search_memory,add_memory,get_project_memories,distill_batch) to all modern agents. - Autonomous Ingest & Distillation: Extracts durable rules and architecture decisions from session logs (OpenAI, Claude, Cursor, Copilot, Antigravity, Pi).
┌─────────────────────────────────────────────────────────────┐
│ SESSION INGEST │
│ OpenAI Exports · Claude JSONL · Cursor · Antigravity · Pi │
└──────────────────────────────┬──────────────────────────────┘
│ ingest catalog / extract
▼
┌─────────────────────────────────────────────────────────────┐
│ STAGING INBOX │
│ Raw captured bullets & noise-filtered facts │
└──────────────────────────────┬──────────────────────────────┘
│ distill_batch / skill
▼
┌─────────────────────────────────────────────────────────────┐
│ LOCAL MEMORY STORE │
│ ~/.agents/memory/USER.md ~/.agents/memory/PROJECTS.md │
│ ~/.agents/memory/concepts/ <repo>/.agents/memory/ │
└──────────────┬───────────────────────────────┬──────────────┘
│ │
▼ ▼
┌─────────────────────────────┐ ┌─────────────────────────────┐
│ IDE INJECTION │ │ MCP SERVER & CLERK │
│ Cursor Rules · Zed Context │ │ search_memory · add_memory │
│ Antigravity AGENTS.md │ │ Universal Tool Integration │
└─────────────────────────────┘ └─────────────────────────────┘
Install directly via pip:
pip install agents-memoryOr install from source in editable mode:
git clone https://github.com/Lolaplex/agents-memory.git
cd agents-memory
pip install -e .# Scaffold initial directories and wire into host IDEs / MCP configs
agents-memory sync --init
# Audit and register local repositories
agents-memory inventoryFill in ~/.agents/memory/USER.md with your profile and stack preferences. Reload your agent to connect to the MCP server.
| Command | Purpose |
|---|---|
agents-memory sync |
Updates canonical AGENTS.md, host rules, and MCP registrations |
agents-memory sync --init |
First-time scaffolding and host discovery |
agents-memory inventory |
Discovers unregistered local repositories across workspace roots |
agents-memory inventory --register SLUG PATH ROLE STACK |
Registers a new repository |
agents-memory ingest catalog |
Indexes local chat transcripts across all supported providers |
agents-memory ingest extract |
Runs heuristic filters to extract durable facts into staging |
agents-memory consolidate |
Ensures no private state leaked into working repository |
- Claude Code: Bound via symlink / canonical
AGENTS.mdand MCP server. - Google Antigravity: Integrated via
.gemini/configrules andagents-memoryMCP. - Cursor: Automatically injects rules and configures
.cursor/mcp.json. - Zed: Configures
context_serversand mirrors assistant skills. - VS Code / Copilot: Ingests session history from local state databases.
The formal, implementation-agnostic layout specification lives in abi/:
abi/WHY.md— Architecture decisions & why Markdown wins over RAG.abi/LAYOUT.md— Directory taxonomy and path contracts.abi/KINDS.md— Typed memory taxonomy (concepts,facts,decisions).abi/MCP.md— Tool definitions and request/response specifications.abi/INGEST.md— Catalog, extract, and distillation pipeline.abi/INJECTION.md— Host rule injection mechanisms.
Run the comprehensive test suite and distillation benchmark:
python tests/run_all_tests.pyMIT License. See LICENSE for details.