Memory that understands time.
The first agent memory system built on a time-aware knowledge graph.
Query what was true last week. Track how claims evolve. Retrieve with confidence — all via MCP.
Most AI agents start every session with a blank slate. Agentic Memory fixes that.
It gives your agents persistent, searchable memory across four domains — code, conversations, research, and git history — stored in a unified Neo4j graph and exposed through the Model Context Protocol (MCP). Agents can recall what they learned yesterday, last week, or last month.
The killer feature? Temporal GraphRAG. Every relationship in the graph carries a validity interval. Ask "what did we decide about the auth flow in March?" and get a temporally consistent answer — not today's guess.
Time-aware graph layer powered by SpacetimeDB. Query what was true at any point in time with deterministic temporal retrieval. Track claim evolution, detect contradictions, and retrieve with confidence intervals.
Structural understanding of files, entities, imports, and on-demand execution tracing — not just text similarity. Your agent knows where things are and how they connect.
Ingest URLs, PDFs, and research reports as searchable findings. Schedule recurring research sessions. Build a living knowledge base that grows over time.
Stores and retrieves past agent/user exchanges by semantic similarity. Never lose context between sessions. Search across months of conversations in milliseconds.
Adds commit/author/file-version history to the same Neo4j graph. Ask temporal questions about your codebase: "who wrote this function and when?"
search_all_memory spans all domains in a single query — code, research, conversations, and git history — with cross-domain relevance ranking.
File watcher automatically updates the code graph as you work. No manual re-indexing required.
Drop-in integration with Claude, Cursor, ChatGPT, Windsurf, Codex, and any MCP-compatible AI. One protocol, every client.
# Install globally (recommended)
pipx install agent-memory-labs
# Or with uv
uv tool install agent-memory-labs
# Initialize in any repository
cd /path/to/your/repo
agent-memory init
# Index your code
agent-memory index
# Start the MCP server
agent-memory serveThat's it. Your repository is now indexed and ready for AI agents.
agent-memory init # Setup wizard
agent-memory status # Repository statistics
agent-memory index # One-time structural ingest
agent-memory index --full # Full rebuild
agent-memory watch # Continuous sync on file changes
agent-memory trace-execution src/app.py:run_checkout --jsonagent-memory research "https://example.com/article"
agent-memory search "What did I read about vector databases?" --domain webagent-memory search "What did we decide about the auth flow?" --domain chatAI Agent (Claude/Cursor/etc.)
|
MCP Protocol
|
+-------+-------+
| Agentic Memory |
+------------------+
| Unified Graph | <-- Neo4j (code + chat + research + git)
| Temporal Layer | <-- SpacetimeDB (validity intervals, PPR)
| Embeddings | <-- Gemini / OpenAI / Nemotron
+------------------+
- Structural code graphs — AST parsing and import analysis, not just file contents
- Unified Neo4j backend — All memory domains live in one graph database
- MCP-native — Primary interface is the Model Context Protocol
- Temporal awareness — Git history and time-sliced queries built into the graph
Agentic Memory is designed to be fully self-hostable:
- Neo4j Community Edition
- Python 3.10+
- Embedding provider API key (Gemini, OpenAI, or Groq)
See docs/SETUP_FULL_STACK.md for detailed instructions.
Works out of the box with:
Claude · Cursor · ChatGPT · Windsurf · Codex
Any MCP-compatible client can connect to agent-memory serve and immediately search across all indexed memory domains.
| Package | Description | Install |
|---|---|---|
agent-memory-labs |
Core Python package | pipx install agent-memory-labs |
agentic-memory-openclaw |
OpenClaw plugin | openclaw plugin install agentic-memory-openclaw |
am-temporal-kg |
Temporal GraphRAG utilities | npm / pnpm |
am-sync-neo4j |
Neo4j sync helpers | npm / pnpm |
We welcome contributions to the core indexing, search, and MCP surfaces.
- Fork the repo
- Create a feature branch
- Run tests:
pytest - Submit a PR
Please keep PRs focused on the self-hostable core.
This project is licensed under the Business Source License 1.1 (BSL 1.1).
- Source available for non-production use, research, and evaluation
- Converts to a standard open-source license after 4 years
- Commercial use requires a license — contact us
See LICENSE for full terms.