Questions
What people ask before they install it
Short answers first, the conditions after. Where a question has a longer answer, it links into the documentation or the full catalogue entry.
What is FluctlightDB?
FluctlightDB is an open-source, embedded memory database for AI agents. An agent writes what happened with experience(), recalls it from a cue with activate(), and makes it durable with checkpoint(). It has a Rust core and a Python-first API, runs inside your process, and stores one brain directory per agent.
How do I give an AI agent long-term memory with FluctlightDB?
Install it with pip install "fluctlightdb[native]", open a brain with connect_embedded("./my-agent-brain"), write memories with brain.experience(text, context=..., salience=...), recall with brain.activate(cue), and call brain.checkpoint() when a unit of work completes. There is no server to start and nothing to configure first.
How do I install FluctlightDB?
Install it with pip install "fluctlightdb[native]". Open a brain with connect_embedded("./my-agent-brain"), write with experience(), recall with activate(), and call checkpoint() so the directory survives a restart. The steps are on /install.
Is FluctlightDB a vector database?
No. A vector database answers "what is nearest?"; FluctlightDB answers "what should this cue bring back?". One activate() call fuses full-text search, vector neighbours and graph spreading activation, and ranks verified memories with a source above ones the agent was merely told. It stores engrams with context, salience and provenance rather than bare vectors.
How is FluctlightDB different from Chroma?
Chroma and Qdrant answer what is nearest, and FluctlightDB answers what this cue should bring back by fusing full-text search, vector neighbours and graph spreading activation inside one activate() call. See /compare/vector-databases.
How is FluctlightDB different from Mem0?
A Mem0-style layer stores facts scraped from chat, and FluctlightDB stores engrams with provenance that the agent records with experience() and recalls with activate(). See /compare/mem0.
Does FluctlightDB need a server or a cloud account?
No. It is embedded: the engine runs in your Python process and a brain is a directory on disk, like SQLite for agent memory. An HTTP server (fluctlight-serve) and a Docker image exist for when you want to reach a brain from another process or language, but they are optional.
Does FluctlightDB need an API key?
No. The canonical sample calls connect_embedded(), experience(), activate() and checkpoint(), and neither that sample nor pip install "fluctlightdb[native]" takes an API key.
Does FluctlightDB do semantic search out of the box?
No. No embedder ships with the engine, so offline recall is lexical: a cue has to share tokens with what was stored. For paraphrase recall, pass a semantic_vector from an embedding model you choose, on both the experience() write and the activate() cue.
Does FluctlightDB work with LangChain, LlamaIndex, the OpenAI Agents SDK or MCP?
Yes. The project ships a LangChain memory and chat history (FluctlightMemory), a LlamaIndex memory (FluctlightLlamaMemory), remember and search handlers for the OpenAI Agents SDK, and an MCP server for Cursor, Claude and Codex, each installed as a pip extra such as fluctlightdb[langchain] or fluctlightdb[mcp]. TypeScript and other languages reach a brain over HTTP (the @fluctlightdb/agent client talks to fluctlight-serve), and there is a Docker image at ghcr.io/voxmastery/fluctlightdb.
Can Claude, Cursor or Codex use FluctlightDB as persistent memory?
Yes, through its MCP server (pip install "fluctlightdb[mcp]"), which exposes tools such as memory_remember and memory_recall. Each assistant session writes to and recalls from a brain directory on your own machine, so memory persists across sessions without sending it to a hosted service.
Which Python versions and platforms does FluctlightDB support?
The native package ships abi3 wheels for Python 3.9 to 3.13 on Linux, macOS and Windows, on x86-64 and arm64. Without the [native] extra you get a pure-Python fallback, which is slower and is not what the measurements were taken on.
Is my agent’s memory safe if the process crashes?
Writes go to a write-ahead log, and checkpoint() writes a whole new generation before atomically advancing a CURRENT pointer, so a process killed mid-write leaves the previous generation intact. Crash recovery is exercised in CI with SIGKILL mid-write and torn log tails. Nothing is durable until checkpoint() returns.
What is the difference between activate() and recall()?
They are two calls with two different return shapes. activate() returns {recalls, active_neurons, hops, myelinated}; recall() returns {hits, lanes_used, mode}. Start with activate() — it is the documented recall path.
How accurate is FluctlightDB’s recall?
On the maintainer’s own frozen July 2026 runs: 72.6% LoCoMo evidence recall at k=5, 96.8% at k=150, and 97.6% LongMemEval-S session recall at k=8. These are maintainer-reported with open harnesses and have not been independently reproduced. An earlier LoCoMo headline was retracted because its protocol counted turns the engine never retrieved.
Is FluctlightDB production-ready?
Not yet, and it says so. It is beta software with one active maintainer, production readiness defaults to false, and there has been no third-party security audit or penetration test. The core API — connect_embedded(), experience(), activate(), checkpoint(), recall() and the on-disk format — is marked stable; everything else is experimental.
How many memories can one brain hold?
Approximate nearest-neighbour search has not been evaluated beyond 100,000 memories. That is not a documented limit, it is an absence of measurement.
Is FluctlightDB free and open source?
Yes. It is dual-licensed MIT or Apache-2.0, the source is on GitHub, and the package is on PyPI as fluctlightdb.
Is FluctlightDB related to Sword Art Online?
No. FluctlightDB is an independent open-source database for AI agent memory. It borrows a word from fiction for its name and is not affiliated with Sword Art Online or its rights holders.
How do I cite FluctlightDB?
Ganesh S, “FluctlightDB: A Memory Model of Data for AI Agents”, 2026. Preprint arXiv:2608.12365, archived at doi:10.5281/zenodo.20949890.
Still deciding?
How it compares with Mem0, Zep, Letta and vector stores, the measurements with their conditions, the preprint and how to cite it, try recall in the browser, or read the six mistakes agents make against this API.