FluctlightDB

Open source · MIT or Apache-2.0 · beta

FluctlightDB, also called Fluctlight An embedded memory
database for AI agents

experience() writes. activate() recalls from a cue. checkpoint() survives the restart. A Rust core, a Python-first API, one brain directory per agent.

pip install "fluctlightdb[native]"

A hosted platform is being built. Until it opens, everything here runs locally: pip install "fluctlightdb[native]"

Open source, MIT or Apache-2.0 · currently beta · source on GitHub · read the paper

An engraving of a human brain seen from the left, the temporal lobe opened to expose the inward curl of the hippocampus and a small almond-shaped body in front of it.
  1. Events Hippocampus
  2. Importance Salience
  3. Focus Executive

Encephalon, lateral view, opened. The structures the engine is named after, marked schematically — this is an engraving, not an anatomical atlas.

What the engine does

Three things, and the third is why it lasts

  1. 01

    Recall by cue

    An engraving of a colossal marble head split open to reveal shelves of books inside, with a figure climbing a ladder within the skull.

    Full-text, vectors and graph spread fused in one activate() call. A memory can come back because of what it sits beside, not only because it shares your wording.

  2. 02

    Evidence outranks chat

    An engraving of a brass balance scale: one pan holds a document closed with a wax seal and hangs low, the other holds only rising smoke and rides high.

    A memory written with a source enters the top provenance tier. When two memories say the same thing, the one that can be checked wins.

  3. 03

    It survives the restart

    An engraving of a sleeping figure whose memories rise as smoke and crystallise into ordered geometric solids above.

    A write-ahead log and atomic checkpoints put the brain on disk. Crash recovery is exercised in CI with SIGKILL mid-write and torn log tails.

Every one of these is documented in the documentation, and the durability tests are in the repository.

Provenance

Written in ink, or written in pencil

An agent hears that the balance is forty dollars, and reads a ledger file that says forty-two ten. Both are memories. Only one is evidence.

Verified

Wallet balance is $42.10

source
file://wallet.json
tier
ledger, verified
rank
1
From chat

“I think I have about forty dollars.”

source
conversation turn
tier
unverified
rank
2

Same cue, wallet balance. The ledger entry is returned first because of where it came from, not because it scored higher. Source: crates/fluctlightdb/src/brain.rs:1848-1861. Opens the repository in a new tab.

Measurements

The number we stand on, and the one we struck

96.8%

LoCoMo evidence recall

at k=150, over 1,982 questions. At the tighter k=5 that a real retrieval turn asks for, it is 72.6%.

99.0%

LoCoMo evidence recall at k=150

withdrawn — a different protocol, which expanded every retrieved turn by its neighbours and then counted neighbours the engine never returned. Plain BM25 reached it too

  • Maintainer-reported
  • Frozen July 2026
  • Harnesses open
  • No independent reproduction yet

All five measurements, their harnesses and the retraction in full · the results in the repository

An engraving of an armillary sphere and a globe held above an open palm, surrounded by radiating beams of light.

Using it

One import, or one container

It is embedded, so there is nothing to run: a brain is a directory, and connect_embedded() opens it in your own process. If you already have an agent framework, there is an adapter for it.

pip install "fluctlightdb[native]"

It plugs into

  • LangChain FluctlightMemory, chat history
  • LlamaIndex FluctlightLlamaMemory
  • OpenAI Agents SDK remember and search handlers
  • MCP Cursor, Claude and Codex
  • Docker ghcr.io/voxmastery/fluctlightdb
  • Node over HTTP from fluctlight-serve

Stable to build on

  • connect_embedded(), connect_agent() and the other connect_* entry points
  • experience(), activate(), checkpoint(), recall()
  • The v4 brain directory on disk
  • POST /api/v1/experience and /api/v1/activate

Getting started · what is wrong with it, six entries, every one sourced.

Contents

On this site

The install line and a working Python sample, the API, the questions, how it sits beside other memory tools, and the preprint.