Lemmalog

Most agent-memory tools bet on 'remembering better': bigger embeddings, smarter chunking, a graph database bolted on the side. lemmalog, built by security researcher Jordy Zomer and released in August 2026, makes a different bet entirely: an agent's memory should be a deductive database. Base facts get asserted at the extraction boundary (an LLM turns conversation into triples), and from there a Rust Datalog engine mechanically derives closures, temporal projections and contradiction candidates; the reasoning happens in a verifiable logic engine, not by re-reading a summary and hoping the model gets the ordering right.

The engine is a stratified, seminaive-evaluation Datalog interpreter with bi-temporal facts (valid_from/valid_to), confidence and provenance carried as semiring annotations, and a why() call that returns a proof tree back to the source conversation turn rather than a black-box score. An MCP server exposes twelve tools to Claude Code or Kimi CLI (lemmalog_observe, lemmalog_query, lemmalog_why, lemmalog_what_if among them), so the host model reads the conversation and asserts facts while lemmalog derives everything downstream deterministically. A hybrid retrieval layer then combines BM25 keyword matching with entity-graph boosting to assemble a token-budgeted context, rather than dumping the whole fact store into every prompt.

It fits long-running, structured work where you want to interrogate what the agent believes and why, investigations, audits, multi-session debugging, more than open-ended chat memory. The why() proof trees and the explicit hypothetical what_if primitive (test a fact against the store, then restore it byte-identically) are the kind of feature that matters specifically when you need to trust and audit a derived answer, not just retrieve a plausible-sounding one. It is a fit for teams already comfortable wiring an MCP server into their agent, not a drop-in for someone who wants zero configuration.

This is a three-week-old project with no tagged release yet; install is cargo build from a checkout, not a package manager. All of the reported numbers (LongMemEval, LoCoMo, the synthetic long-horizon eval) are single or triple-repeated runs the project ran itself, and the README is unusually candid about measurement noise: on one benchmark, an answer 'flipped 1.00 to 0.00 across two same-configuration runs' because the model used rejects a temperature parameter. Fact extraction still depends entirely on an external LLM call to turn prose into triples; the Datalog engine reasons over facts, it does not read English itself, so extraction quality is a hard ceiling on everything downstream.

lemmalog is worth watching if deterministic, provenance-tracked reasoning over agent memory, rather than another vector-similarity layer, is what you actually need, and if you are comfortable running pre-release Rust from source. The engineering rigor (differential testing against a brute-force oracle, honest reporting of reverted experiments) is well above what a three-week-old repo usually shows. But there is no packaged release, a small contributor base, and every benchmark number is the project's own single-digit-run measurement; treat it as a serious research prototype worth tracking, not yet a production dependency.