Tgms MCP server

TGMS — Agent-Native Bi-Temporal Graph Management System

-97%25-brightgreen.svg)

A temporal graph database whose query surface is built for LLM agents — and whose answers can be audited claim by claim.

Project page & blog: https://zxf-work.github.io/tgms/ · Paper: paper/main.pdf

LLM agents are unreliable at exactly the things temporal graph analytics requires: arithmetic, identifiers, and asserting only what the evidence shows. TGMS's answer is architectural — give the model no opportunity to do any of them:

distinguishes evolution ("the edge ended") from correction ("we were wrong"), so agents can answer "what did we believe on March 1?" — a question latest-state snapshots and the RAG configurations we evaluated cannot express. Bi-temporality itself is inherited, not invented here — it has a four-decade literature, a place in SQL:2011, and production databases built around it. We measure against the clearest of those, XTDB: fed the same operation stream, the two systems agree on believed state at 400 of 400 probe points, with TGMS 3.9–4.7× faster at correction-heavy ingest on 23–27× less disk (the head-to-head);

• a bi-temporal property graph (valid time × transaction time) that

answer, TGMS now tells you when that happened: tgms trace check reads a saved answer's dependency scope against the event log — no recompute, no store lock required — and returns FRESH / POSSIBLYSTALE / UNDECIDABLE, sound in the direction that matters (it never calls a stale answer fresh). Measured across two injection campaigns, 6,978 trials: 0 false-fresh verdicts on 447 changed answers in the M4 campaign (run 2, the record of account) and 0 more on 1,852 changed answers in the M5 carve-2 campaign (28,044 trials), where the obvious cheap check — "did the correction touch a row in the stored result?" — is wrong on 47.4%** of the M4 trials;

• and because a belief can be corrected after you've already acted on an

itself: tgms artifact register/check/refresh turns it into a named, generation-numbered artifact — refresh recomputes only what you ask, the old generation stays byte-identical on disk, and a refresh propagates one hop to whatever else was built on top of it, even when that dependent's own scope was never touched. Measured across the M5 maintenance campaign: 0 false-fresh in 37,371 trials, 0 false-safe over 5,867 propagation decisions (99.0% resolved without recomputing anything), and a 600/600 pinned-answer exemption;

• and a saved result you want to keep, not just check, can now maintain

paths, δ-motifs, snapshot diffs, burst detection, interval joins, grouped aggregation over edge events, and the belief log itself) — typed, deterministic, bounded, cost-guarded, exposed as tools (MCP or in-process); identifiers must come from a resolver, arithmetic from a compute operator;

• 15 verified temporal operators (reachability over time-respecting

plans are statically validated (including a grounding rule that makes fabricated identifiers impossible and output-field contracts that reject invented result paths), executed deterministically with content-addressed traces, and every claim in the written answer is machine-checked against the trace that produced it — including truncation taint, so "correct arithmetic over incomplete evidence" is caught too;

• a Planner–Executor–Verifier loop: the LLM only plans and reports;

columnar

• a purpose-built native storage engine (Rust, PyO3): bi-temporal

Shortened. The full README is on GitHub.