Syntagraph is an evidence-backed hypergraph for agents. Every fact is checked against its source before it is written, and every answer comes back with its exact size.
→ write Claim C8 quote "typed corpus plans outperform RLM on corpus-level queries" source P031 by seed:v14 rung 1 quotes exist verbatim ✗ quote not found in P031REFUSED · nothing was recorded
→ verify_claim C5 rung 1 quotes exist verbatim ✓ found at 690–737 in P026 rung 2 numbers match their sources ✗ not in any cited source: 5 passed 1 of 2HELD · the quote is real, the number is not sourced
SELECT( STAR(H10), where: { type: "Claim" } ) sort edges cardinality 1 exact true truncated false C7 No prior work trains a cost-bounded recursive policy for corpus-level aggregation. origin: hypothesisedEXACT · all 1 of 1, in 11 ms
→ item H6 hypothesis refute if operation features add less than 2 points AUC result added 0.9 spent $96 of $120 status refuted by its own stop rule kept in STAR(P174), STAR(O6) so the next agent sees it before trying againRECORDED · as a negative result
Agents cannot write what they cannot quote. The check runs before the fact lands.
Counts, absences and comparisons are queries. You are told what was left out.
Ask what was true then, and separately, what was known then.
Memory layers are built to remember more. Nothing in them asks whether the thing being remembered is true. An agent stores a confident sentence with a citation that does not say that, and from then on it is "context".
Others help an agent recall. Syntagraph makes what it records defensible, countable and re-runnable.
Each write climbs a ladder. Rung 1 re-slices every quote from the stored source and confirms it is there word for word. Rung 2 checks every number. Fail a rung and nothing is recorded.
rung 1 quotes exist verbatim ✓
rung 2 numbers match their sources ✗ → held
rung 3 an independent model, shown only the
cited sentence, agrees · optional
A query starts from an entity's star: every fact it takes part in. When results are ranked for a large hub, you still get the full count.
Top-k tells you what it found. This tells you what there is.
The store keeps when something held in the world and when the system learned it. You can replay what the program believed on the day a plan was locked.
A hypothesis carries the result that would refute it and the budget it may spend. When the rule fires, it stops itself and the negative result stays findable.
Recursion stops at two levels. Splits are subtrees of a typed expression and merges are exact algebra, so any conclusion can be re-run when new evidence arrives.
This is our own research program, read straight from the graph. Decisions wait for a person. Checks, refusals and stopped ideas are all on the record.
Drawn with this page's own styling from the live tenant's overview on 3 October 2026. It is not a screenshot of the app.
Documents are stored whole, so any quote can be re-sliced from the original text later.
Relations with any number of participants, each playing a named role, each linked to its evidence.
STAR, SELECT, MEET, FACET and seven more. Agents compose them into typed expressions.
One tool contract over MCP: scan, read around, trace, compare, verify, propose a decision.
Eight rungs between a draft and a recorded fact. Proposals become approval cards for a person.
The star of an entity is every fact it takes part in. It is complete by construction, which is why a count over it is exact and why a missing fact shows up as missing.
Memory layers, search engines and graph databases are good at what they are for. This is what each kind of tool is designed around.
| Memory layers | Search engines | Graph databases | Syntagraph | |
|---|---|---|---|---|
| Designed to | Remember across sessions | Find the most relevant passages | Model and traverse relations | Record only what holds up |
| A write is checked against its source | Not the pitch | Not the pitch | Schema only | Before it lands |
| An answer states its full size | Top results | Top results | Yes, for a query | Always, even when ranked |
| What was known on a past date | Some keep fact history | No | If you model it | Built in |
| Hypotheses, decisions, stop rules | No | No | If you model it | First-class objects |
| Speed at billions of rows | Varies | Their strength | Varies | Not our claim |
Based on how products in each category describe themselves on their own sites, October 2026. We have not benchmarked them here.
We are building 300 test questions with exactly known answers to compare five retrieval methods at equal cost. 212 are ready. When the comparison runs we will publish the questions, the method and the cost per method, so anyone can re-run it.
Syntagraph is in private preview. It runs as an MCP server and a Claude Code plugin, next to the tools your agents already use.