EngramRAG: ontology memory for long-term human–AI work

In long-horizon work, the bottleneck is what the model gets to read

Frontier language models can reason remarkably well inside a conversation. Long-term work asks for something a model cannot supply on its own: intent, constraints, prior decisions, evidence, and later corrections must still be available when the next session begins.

A larger context window increases capacity. It does not decide what should persist, which record is current, or which facts belong together. People cannot reconstruct every piece of tacit context in each new prompt, and even the strongest model cannot recover evidence it never received.

We call this the shared-context externalization bottleneck.

Reliable human–AI work depends not only on model capability, but on whether knowledge exists as persistent, reviewable, searchable external state.

EngramRAG is Consilience’s answer to that problem. Human-editable Markdown remains the place where knowledge is written and owned. A derived ontology becomes a machine-navigable index of identities, relationships, provenance, and revision. Agents search, read, and write through the same document system, so verified work can become useful context for the next task without allowing a machine interpretation to replace its source.

One memory, three trust boundaries

EngramRAG separates memory by authority instead of pouring everything into one vector store.

That separation is also a safety boundary. Claims about the world belong in reviewable documents. Preferences about how to work belong in procedural memory. Content read from a web page or screen cannot silently promote itself into a durable instruction.

When an agent learns something, it does not train model parameters or insert an unreviewed fact directly into the graph. It checks for duplicates and writes an ordinary Markdown note. That note earns authority through the same review and extraction cycle as every other document.

A lifecycle that returns to the document

EngramRAG carries knowledge through a complete memory loop:

1. Record and import. People and agents write Markdown. Imported files are converted while their originals remain available.
2. Extract incrementally. Only changed documents are processed for entities, relationships, and statements.
3. Bind every relation to a source. Each document owns a named graph. Re-extraction replaces that graph atomically, so triples from deleted text do not linger.
4. Resolve identity without destructive merging. Surface forms, embeddings, types, document evidence, and model judgment propose whether mentions refer to the same subject. Same, broader, narrower, related, and distinct decisions remain reversible overlays.
5. Retrieve and assemble context. Query signals seed the graph, Personalized PageRank carries relevance through relationships, and source passages plus relationship evidence are assembled within a bounded context.
6. Act with controls. Agents answer and edit through the same search, read, and write paths, with permissions, diffs, checkpoints, and rewind boundaries.
7. Return knowledge to the cycle. A verified finding or corrected decision becomes Markdown again. The next extraction updates the graph beneath it.

The goal is not a graph that grows forever. It is a memory that can return to its source, accept a human correction, and rebuild the machine-readable structure around that correction.

Search follows meaning and structure

Ordinary RAG asks which passages look most similar to a question. EngramRAG asks what the question is connected to.

Its explicit search path combines four signal families:

• lexical matches to entity names;
• semantic similarity to entity-label embeddings;
• similarity to embeddings of extracted relation triples; and
• document and chunk retrieval through dense embeddings and BM25.

These signals seed the graph. Personalized PageRank spreads relevance across relations, aliases, hierarchies, and document mentions. Context assembly also reserves space for bridge relations: a lower-ranked intermediate edge can be the piece that explains how two highly ranked facts connect.

Automatic turn context stays deliberately small. A lightweight lexical-and-graph path contributes at most eight short results and marks them as partial or potentially stale. Deeper questions can invoke explicit search and graph traversal to verify the evidence.

The same model performs differently with a better memory path

We test EngramRAG by removing parts of the system while holding the model, task, corpus, or production path fixed. The strongest results are not model comparisons. They show what changed when the surrounding memory structure changed.

The code ablation is the clearest signal. Text search solved 14 of 21 tasks. Adding embeddings alone still solved 14. Adding the relation graph raised the result to 18; the shipped graph path reached 19. In a separate hard run, the same comparison was 6/12 versus 12/12.

The surrounding system also changed the economics of the agent loop. In the balanced run, cost per successful location fell from $0.180 to $0.092, latency from 22.0 to 17.5 seconds, and tool calls from 7.76 to 5.29. These are scoped live-run measurements, but the conclusion is direct: in this experiment the gain came from relation structure and a path that made the agent use it—not from swapping in a stronger model or adding embeddings alone.

The retrieval result repeats across documents. Full fusion led keyword-only search by 30.7 and 38.7 percentage points in two real vaults. On MuSiQue, removing graph ranking reduced the share of questions with every required passage in the top ten from 63% to 25%. The model did not become more capable between arms; the system became better at selecting what the model could read.

Read the retrieval evaluation.

Memory begins by reading the source faithfully

The memory graph can only be as useful as the material that enters it. Consilience converts PDFs, office documents, spreadsheets, slides, web pages, images, and other sources into editable Markdown while retaining the original. Tables, headings, reading order, and source boundaries are preserved so the ontology is built from a faithful record rather than a lossy summary.

We measured that entry point again on 31 August 2026. A single locked run of the Consilience parsing pipeline, powered by Gemini 3.6 Flash, completed all 200 released DP-Bench documents without a failed prediction.

The candidate exceeded the then-listed best TEDS by 0.28 points and TEDS-S by 0.79 points. It used one full-corpus run—no best-of-N selection, ground-truth lookup, manual correction, or per-document repair. Official listing remains under review, so this is a reproducible submission candidate rather than a claimed official rank. See the DP-Bench public leaderboard and submission record.

Consilience built the parsing, semantic-serialization, normalization, and evaluation pipeline. It did not train Gemini 3.6 Flash; the foundation model was provided by Google through an external provider.

Provenance, not truth

A model extracting a triple from a sentence does not make the triple true. EngramRAG makes that distinction structural.

• Every extracted relation remains attached to a document-specific named graph.
• Human assertions and model-proposed analysis live in separate layers.
• Identity resolution is an overlay, so a mistaken merge can be reversed and a user’s distinct decision can block a future automatic merge.
• Later decisions can be proposed as revisions of earlier ones without erasing the earlier assertion.
• Types can veto incompatible merges, while the source document remains available for inspection.

The operating principle is provenance, not truth. EngramRAG preserves why a relationship exists and who authored, extracted, or approved it. The ontology helps people and agents navigate knowledge; it does not declare what they must believe.

From one query to the full memory lifecycle

EngramRAG changes the unit of RAG. The unit is no longer one query and the passages nearest to it. It is the full lifecycle through which knowledge is recorded, connected, retrieved, corrected, and used again.

People should not have to reconstruct a perfect prompt for every session. Models should not have to start from a handful of disconnected files. Consilience places human-owned, editable documents between them and builds a reversible relation index underneath. The source stays readable. The structure stays navigable. Corrections survive the conversation in which they were made.

The measurements point in one direction: with the model held constant, memory structure changes what an agent can find, how reliably it completes the task, and what a correct result costs. Model capability and memory architecture multiply each other.

EngramRAG is software developed by Consilience as part of its parsing, memory, and retrieval pipeline. It can use externally provided foundation models during extraction and reasoning; it is not itself a foundation model, and Consilience did not train those third-party models.