The memory terrain
for AI agents.

An embedded property graph where facts supersede rather than overwrite. Every supported read is scoped. Vector recall is deterministic and bounded to a subgraph. One file, one process, and no model is ever called from inside the database.

cargo add topodb@0.1.0

0.1.0 — breaking public API changes wait for 0.2.0. The on-disk format is versioned and migrates in place.

Or skip the Rust: the Claude Code and Cursor plugins run the same engine over MCP — recall injected at session start, memory captured as you work.

Read the source →

Benchmarks — LongMemEval-S

Recall is the job. Here is recall, measured.

retrieval leg

R@1

R@3

R@5

R@10

text (BM25)

0.872

0.932

0.953

0.979

vector (MiniLM)

0.864

0.953

0.977

0.989

hybrid (RRF)

0.894

0.966

0.987

0.996

A hit means a gold-evidence session appears among the first k retrieved sessions. Hybrid retrieval puts the gold evidence in the top-5 98.7% of the time across the full 500 questions — and leads every leg at R@1.

The honest split: end-to-end QA accuracy is 63.3% (stratified 90 questions, gpt-4o reader and the official LongMemEval judges). That gap is reader-stage, not memory — the same retrieval had already placed the evidence in the reader's context ~99% of the time. A memory layer is measured on whether it surfaces the right memory; what an LLM does with it is measured separately.

Ingesting each conversational turn as its own memory — rather than whole sessions — lifted R@1 by up to +0.12 for free: deterministic, no model in the loop. It is the harness default.

LongMemEval-S · 500 questions (470 graded, 30 abstention)
session-level Recall@k · turn granularity
all-MiniLM-L6-v2, host-computed · measured 2026-08-14
benchmarks/longmemeval · python -m lme.run

MiniLM is deliberately small — it isolates the engine's ranking from embedder quality. With text-embedding-3-large, retrieval@5 reached ~1.00 on a balanced sample. The QA figure depends on an external reader and judge, and moves when they do.

The memory terrainfor AI agents.

Recall is the job. Here is recall, measured.

The memory terrain
for AI agents.