A memory substrate under measurement
We set out to build AI memory the way biological brains remember. Then we tested our own claims with pre-registered experiments—and we publish what the benchmarks say, including when they say no.
Most AI marketing describes what a system is meant to do. This page describes what ours has been shown to do. Every experiment below was pre-registered: the pass/fail criteria were frozen before any result existed, and when a claim failed its test, we closed the claim instead of moving the goalposts.
A memory system you would trust with your organization's knowledge should be built by people who report their negative results. Here are ours.
Summer 2026: a sequence of pre-registered experiments on graph-based, biologically inspired retrieval, run against strong embedding-similarity baselines on a public benchmark and on real incident postmortems.
Our founding hypothesis was that relevance is consequence, not similarity—that a graph of causal connections would out-retrieve embedding search. It didn't. On a public long-conversation benchmark, similarity reached 86.0% recall@200 to our typed graph's 77.6%. On our own consequence-question probe over real postmortems—conditions chosen to favor the graph—similarity found the author's root cause at median rank 28; the graph, rank 77.
Pre-registered falsification clause applied as written. The claim is closed, not rephrased.
Our extraction pipeline does find genuine causal structure in real documents—on eight incident postmortems it produced hundreds of typed cause/effect relations, most touching the true root-cause passages. But deleting every one of those edges changed retrieval quality by approximately nothing. Causal structure is valuable; using it to rank search results is not where the value lives.
Measured by ablation: identical corpus, identical queries, edges removed, ranks essentially unchanged.
The sharpest finding cut both ways: on "what caused this?" questions, every mechanism we tested—ours and the industry-standard baseline—ranked a topically adjacent passage from the same document above the actual causal answer, on every single question. Similarity wins the comparison and still doesn't solve the problem.
Consequence questions are real, and they are unserved. That finding now drives our roadmap.
The engineering substrate passed every integrity gate we threw at it: bit-identical replay determinism across full experiment reruns, write-ahead-logged durability with crash recovery, complete provenance from every derived insight back to its sources, and honest instrumentation that caught our own pipeline defects before results did.
Trustworthy memory infrastructure turned out to be the part worth keeping.
A research substrate for accountable AI memory—the parts our measurements kept.
Write-ahead logging, atomic snapshots, crash recovery, and a deterministic-replay architecture: the same inputs always rebuild the same memory.
Document ingestion that extracts who-caused-what as typed, provenance-stamped records—kept for explanation and audit, not for ranking.
Every derived record traces to its sources. Nothing enters memory without a recorded origin; nothing is silently rewritten.
Namespace memories by persona, tenant, or domain for multi-context use.
Instrumentation built for experiments: write-side witnesses, edge censuses, and integrity checks that fail loudly.
Docker-native deployment with a fully typed, async-first client.
Three directions, each gated on its own pre-registered experiment. None of these is a capability claim; each becomes one only if it passes.
The surviving form of our thesis: instead of causal edges at query time, causal annotations written into the representation itself—so that "near" comes to mean "shares consequences," and ordinary similarity search exploits it. The kill test is designed; the prior is honestly unfavorable; we'll publish either way.
Since no ranking function isolates a causal answer, synthesis has to— under a contract: every claim in a generated answer must cite a retrieved memory, and the citation must survive an automated check. Deterministic retrieval, accountable synthesis.
Parked intentions, deferred decisions, and wake conditions evaluated deterministically as events arrive—prospective memory, aimed at a measured weakness of today's frontier models on deferred-intention tasks.
Pass/fail criteria are frozen before results exist. Thresholds are never adjusted after a number is seen.
Five falsification clauses have fired in this program. All five were honored as written.
We don't demo, sell, or license capabilities that haven't been validated end to end. When something passes its gates, you'll see it here first.
We write up what we build and what we measure—including the experiments that fail. If your organization cares about memory it can audit, we'd like to hear what questions you need answered.
enterprise@engramforge.com
This page describes a research program, not a generally available product. Last revised August 2026 to reflect measured results, including negative ones. We are not taking demo or sales conversations for unvalidated capabilities.