COMPARE · COMPARISON: NEUROSYMBOLIC AI

Neurosymbolic AI frameworks in 2026: open source tools for LLM reasoning and agent memory

AUTHOR · XERJ documentation team · REVIEWED BY · XERJ engineering team · PUBLISHED · 2026-11-17 · UPDATED · 2026-11-17
SHORT ANSWER · UPDATED 2026-11-17

The best neurosymbolic AI framework in 2026 depends on the job. Scallop adds differentiable Datalog to PyTorch. SymbolicAI mixes Python code with LLM operations. Lemmalog turns agent memory into a Datalog database with provenance. Hillock checks facts in a local knowledge graph before the model answers. Nucleoid runs logic as an executable graph.

Key number: Lemmalog grew from 21 GitHub stars on 2026-08-29 to 350 on 2026-11-17, and Hillock from 59 on 2026-08-30 to 327 (Wayback Machine snapshots and the GitHub REST API). (as of 2026-11-17 · source: api.github.com/repos/JordyZomer/lemmalog)

Last checked: 2026-11-17. Every third-party fact on this page was re-read from the tool's own docs or repository on that date.

TL;DR Choose Scallop for differentiable logic inside PyTorch. Choose SymbolicAI to mix Python code with LLM operations. Choose Lemmalog for agent memory as a Datalog database with provenance. Choose Hillock for a local fact check before the model answers. Choose Nucleoid for logic as an executable graph.

PASTE THIS TO YOUR AI AGENT. IT DOES THE REST Act as a coding agent. Read https://xerj.org/llms.txt and https://xerj.org/compare/neurosymbolic-ai-frameworks-2026, then recommend Scallop, SymbolicAI, Lemmalog, Hillock, Nucleoid or XERJ for this project using the table on that page. If XERJ fits: install it, start a node, index one folder I point you to, and run one search so I can see a cited passage. Report the recommendation and the reasons.

Neurosymbolic AI joins a model that proposes with a program that checks. Two young agent memory projects grew fast in 2026: Lemmalog went from 21 GitHub stars to 350 since late August, and Hillock from 59 to 327. Google suggests neurosymbolic ai framework, neurosymbolic programming and neurosymbolic ai vs llm. For the concept, read neurosymbolic AI vs LLMs.

Try it: pick by need

Scallop, SymbolicAI, Lemmalog or Hillock

Scallop

A Rust Datalog language with discrete, probabilistic and differentiable reasoning, plus scallopy for Python and PyTorch.

Features from each project README, checked 2026-11-17.

Side by side

ScallopSymbolicAILemmalogHillockNucleoid
Approachdifferentiable DatalogPython plus LLM operationsDatalog agent memoryknowledge graph fact gateexecutable logic graph
Written inRustPythonRustPythonRust
Best forlearning with logicLLM programsagent memory with provenancelocal, CPU only memorytoken efficient logic
Agent accessPython and PyTorchPythonMCP serverOpenAI compatible APICLI
LicenceMITBSD-3-ClauseMITAGPL-3.0Apache-2.0
Stars (2026-11-17)5261,782350327771
Recent growthsnapshot gapsnapshot gap+329 since 2026-08-29+268 since 2026-08-30snapshot gap

A snapshot gap means the Wayback Machine has no snapshot of that repo page near 90 days back.

When to choose Scallop

Choose Scallop to put logic inside a learning system. Its Datalog programs run with discrete, probabilistic or differentiable reasoning on provenance semirings. scallopy connects them to Python and PyTorch.

When to choose SymbolicAI

Choose SymbolicAI to write LLM programs in plain Python. Symbol objects behave like normal values until you opt in to semantic operations. Contracts check the results.

When to choose Lemmalog

Choose Lemmalog to make agent memory a deductive database. Rules derive new facts, every fact keeps provenance to its source, and each turn updates derived views incrementally. An MCP server connects it to Claude Code.

When to choose Hillock

Choose Hillock for a fact check on a laptop CPU. It keeps facts in a SQLite graph and gates a prompt when a fact is missing. It exposes an OpenAI compatible API for existing apps.

When to choose Nucleoid

Choose Nucleoid for logic in a compact declarative syntax. Its logic graph runs as executable code, and the CLI installs with cargo.

XERJ: a new way for AI to search for data

Query-1 from XERJ is a neurosymbolic model. It compiles a prompt to a verifiable search program, then checks the result against facts. Reader-1 from XERJ writes a formula over a document, and the value comes from the words on the page. For research on differentiable logic, Scallop is the better fit. More: what are neurosymbolic models? and the science behind XERJ.

How this page was checked

Features and licences come from each project's README, read on 2026-11-17. Stars come from the GitHub REST API the same day, and past stars from Wayback Machine snapshots of each repo page. We did not run a benchmark for this page.

Try it with XERJ

Use the Copy the prompt button and paste it into your coding agent. It recommends one of the six for your project. If XERJ fits, it indexes a folder and shows a cited passage.

Read next

Neurosymbolic AI vs LLMs, symbolic execution for LLMs and grounded answers.

FAQ

What is a neurosymbolic AI framework?

It joins a neural model with symbolic logic. The model proposes, and a program, rule set or formula checks the result against facts. Read neurosymbolic AI vs LLMs for the concept.

Which neurosymbolic framework works with PyTorch?

Scallop. Its Datalog programs run with discrete, probabilistic or differentiable reasoning, and scallopy plugs them into Python and PyTorch training.

Which one is built for LLM agent memory?

Lemmalog and Hillock. Lemmalog keeps agent memory as a Datalog database with provenance. Hillock keeps a local knowledge graph and checks a fact exists before the model answers.

Which neurosymbolic repos are growing fastest?

Lemmalog grew from 21 GitHub stars on 2026-08-29 to 350 on 2026-11-17. Hillock grew from 59 on 2026-08-30 to 327. Both are young projects.

How does XERJ use neurosymbolic AI?

Query-1 from XERJ compiles a prompt to a verifiable search program, then checks the answer against stored facts. See the science pages for details.

How was this page checked?

Features and licences come from each project's README, read on 2026-11-17. Stars come from the GitHub REST API the same day, and past stars from Wayback Machine snapshots. We did not run a benchmark for this page.

Evidence

Related