Splits memory into three tiers: working (in-session, ephemeral), episodic (graph-based conversational history in Neo4j), and profile (long-term user facts in Postgres). Agents interact through a REST API, Python or TypeScript SDKs, or a native MCP server. An optional retrieval agent layer (split-query and chain-of-query strategies) orchestrates multi-tool retrieval for complex questions rather than issuing a single lookup. LLM-agnostic across OpenAI, Anthropic, Bedrock, and Ollama, and ships an in-repo evaluation harness covering LoCoMo, WikiMultiHop, and HotpotQA.
- Storage
- Neo4j 5.23 with APOC and GDS plugins for episodic graph memory, PostgreSQL with pgvector for profile memory, and an ephemeral in-process store for working memory — stood up together by the official docker-compose.yml.
- Retrieval
- Graph traversal and search over the episodic Neo4j store, SQL plus vector lookup over profile facts, and an optional retrieval agent that decomposes a complex query into multiple tool calls before synthesizing an answer.
- Self-host
- Self-host: moderate
- License
- Apache-2.0
- Pricing
- Core is free and Apache-2.0, self-hostable via Docker Compose at no cost. A hosted MemMachine Cloud offers free-tier API keys; the vendor states an enterprise version with additional features and dedicated support is 'available soon' with no published price as of 2026-07-25. · Free + paid
- GitHub stars
- 3,341
- Last release
- 2026-05-18
- Last commit
- 2026-07-20
- First catalogued
- 2026-07-25
Strengths
- Three explicit memory tiers mapped to genuinely different latency and durability needs, rather than one undifferentiated store
- Nine first-party framework adapters plus native MCP over both stdio and HTTP — unusually broad for an Apache-2.0 project
- LLM-agnostic across OpenAI, Anthropic, Bedrock, and Ollama, so the backbone is not locked in
- Ships a real in-repo evaluation harness (LoCoMo, WikiMultiHop, HotpotQA) rather than only publishing marketing numbers
- Retrieval-agent layer decomposes multi-hop questions instead of relying on a single similarity lookup
Watch out
- Self-hosting means running Postgres and Neo4j and the app container — three services — despite the '5 lines of code' framing; the quickstart itself notes it requires a running MemMachine server
- The promised enterprise tier has no published price or date, so total cost of ownership at scale is unknowable today
- The headline 91.69% LoCoMo and 93.0% LongMemEval-S figures are self-reported in a MemMachine-authored preprint (arXiv:2604.04853), not independently reproduced
- Backed by MemVerge, a company pivoting into agent memory from memory-tiering hardware — worth watching for roadmap continuity
- A sizeable open-issue backlog relative to project age; evaluate before production use
Best for
- Teams that want separate working, episodic, and profile tiers and are comfortable operating Postgres plus Neo4j
- Multi-framework shops wanting one memory backend wired into LangChain, CrewAI, LlamaIndex, or n8n via first-party adapters
- Applications where multi-hop reasoning over conversation history matters more than single-shot recall
How it integrates
Benchmark results
No sourced results yet.
Sources
- Repo metadata — 3,341 stars, Apache-2.0, main branch last commit 2026-07-20 (vendor)
- Confirms the three-service self-host stack: Postgres/pgvector, Neo4j 5.23 with APOC+GDS, and the app container (vendor)
- MemMachine paper — self-reported LoCoMo 91.69% (gpt-4.1-mini backbone) and LongMemEval-S 93.0% (paper)
- Vendor site — hosted cloud with free API keys; enterprise tier 'available soon', no published price (vendor)
- Python client SDK on PyPI, v0.3.9 (vendor)
- In-repo integration folders confirming the nine first-party framework adapters (vendor)
Last verified 2026-07-25 · updated by discover-frameworks