Open source memory for agentic AI
Give AI systems memory that lasts.
HumemAI is an open source organization. We publish research, code, and benchmarks for a memory layer that stays persistent across sessions, adapts to mixed data types, and remains inspectable instead of turning into a black box.
Why memory
Most agents still behave like stateless interfaces with better wording.
HumemAI focuses on what should persist beyond a prompt: what happened, what matters now, and how structured knowledge should stay available over time. That means treating memory as a real system layer instead of a side effect hidden in context windows.
The result is a stack that can hold documents, tables, graphs, and traces in forms that remain inspectable, replayable, and useful to both people and agents.
Open source
Everything we make is free to use, modify, and redistribute.
The code, Python packages, benchmarks, and papers are published under open licenses.
Contribute
Built in the open, with anyone who wants to help.
Issues, pull requests, benchmark reports, and research collaborations are all welcome. Pick a piece of software and start from its repository on GitHub.
Software
Open source software.
Databases and the tools around them.
ArcadeDB
A database that keeps documents, graphs, vectors, and time series in one engine. Run the official server, or run it inside your Python process with uv.
Audit-Ready Memory
A public-interest project building local-first, inspectable memory infrastructure for AI systems that need audit trails, deterministic replay, and explicit deletion.
CypherGLOT
A Cypher-to-SQL compiler that lets developers think in graph terms, then lowers admitted Cypher through an internal AST and IR into backend-aware SQL for host runtimes.
HumemDB
A Python layer over SQLite, DuckDB, CypherGLOT and LanceDB. It routes table, graph and vector work to the embedded engine that suits it.
Research
Research on AI memory and databases.
Human-like memory for machines, multi-model databases, and benchmarks.
ArcadeDB
ArcadeDB keeps documents, graphs, vectors and time series in one engine. How it compares with the specialist engines for each kind of data, with every setting published.
DBBench
An independent, open benchmark of database engines. The same tasks on the same machine, with published settings and every result tied to the exact engine version.
ArcadeDB in Python
One embedded database for documents, a graph and vectors, running inside the Python process. How it works, how it compares with the specialists, and what it costs.
RoomKG Meta-Policies
A neuro-symbolic meta-policy that learns which symbolic memory heuristic to apply for answering, exploring, and forgetting, while keeping execution symbolic.
KG Memory Transfer
A learned policy that decides which short-term memories to keep in long-term knowledge-graph memory, for an agent that only sees part of its world.
RoomKG Baselines
A benchmark for temporal knowledge-graph memory when the agent only sees part of its world, with symbolic and neural baselines to compare against.
News
Latest news from HumemAI.
Recent writing, releases, research, and benchmark updates from HumemAI.
