FluctlightDB
The memory engine for AI agents — not a vector database with an agent SDK bolted on.
Mission
Goal: become the default database for agent memory — the way SQLite became the default embedded DB for apps.
Long-term agent memory is a third data model (alongside relational facts and vector similarity). FluctlightDB defines engine-level experience() / activate() semantics — episodes, cue-driven recall, provenance, consolidation — not app glue on top of Chroma or Mem0.
Who it's for — agents that learn and retain across sessions (from chat, tools, files, and APIs), recall under paraphrase, prefer verified evidence over casual conversation at recall time, and — in monorepos — share a project brain across Cursor, Claude Code, and Codex (fluctlight-project init). No hosted SaaS required.
What “learning” means — not fine-tuning. The agent writes episodes (experience()), recalls under new cues (activate()), and consolidates over time (sleep() / checkpoint()). See the full README and Manifesto.
Typical fits: coding agents, ops bots, research assistants, NPCs.
Install
pip install "fluctlightdb[native]"
from fluctlightdb import connect
brain = connect("/tmp/my-agent-brain")
brain.experience("User prefers dark mode", context="settings", salience=0.8)
print(brain.activate("dark mode")) # offline lexical cue (paraphrase needs semantic_vector)
brain.checkpoint()
HTTP-only (no Rust extension): pip install fluctlightdb
Benchmarks (June 2026)
| Benchmark | Metric | Result |
|---|---|---|
| LoCoMo (10 conv, 1,982 q) | Mean evidence recall @ k=150 | 96.8% MiniLM / 97.0% mpnet (no expansion; @5=72.6%/75.1%) |
| BEIR SciFact | nDCG@10 (index mode) | 0.645 (ties Chroma + MiniLM) |
| FAMB | Macro (index / agent) | 98% / 97% |
Frozen JSON: benchmarks/results/paper-2026-07-09.json
LoCoMo evidence recall ≠ Mem0 LLM-as-judge QA — different metrics; compare only when labeled. (The old 99.0% headline was ±3 neighbor-expansion inflation, now deprecated.)
Docs
- Getting started
- Multi-agent monorepos —
fluctlight-project init, MCP, hooks, handoffs - Platform compatibility — Windows, macOS, Linux
- Full README & reproduction
- Paper (LaTeX)
1 maintainer
Metadata
1 maintainer
Release files for fluctlightdb 0.5.20
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | |
|---|---|---|---|
| fluctlightdb-0.5.20.tar.gz | 54.9 kB | Details |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Reset |
|---|---|---|---|---|
| fluctlightdb-0.5.20-py3-none-any.whl | Python 3 | none | any | Details |
Total release size: 121.6 kB
Release files / fluctlightdb-0.5.20.tar.gz
| Download URL | fluctlightdb-0.5.20.tar.gz |
|---|---|
| Size | 54.9 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
9216003b2ed37a26d0ab07ce744c0380833873336dafc9457063fd8890d9f149
|
|
BLAKE2b-256 checksum How to use checksums |
a90a6862126061bf1bec90490200679a82633a932aa43272c9fe0fc587dde3fe
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Release files / fluctlightdb-0.5.20-py3-none-any.whl
| Download URL | fluctlightdb-0.5.20-py3-none-any.whl |
|---|---|
| Size | 66.7 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
434c09b4328438fcb7cd9e5f1030ab04f0c6a0c4d1b93565707712fc77763fde
|
|
BLAKE2b-256 checksum How to use checksums |
80e7c937802297c2e3903b1026ede5518ab787b0d85c046070842d354ca220ce
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
