Agents that can answer as of the date of the dispute
A contract question is almost never about today. It is about what the agreement said, and who was authorised to sign it, on a date that has already passed.
The current version cannot answer a historical question
Legal work is bi-temporal by nature and almost every document system is not. That mismatch is the whole failure.
An amendment is a new file, not a change
The executed agreement and its three amendments sit in the repository as four separate documents. Nothing records that clause 7.2 in one replaced clause 7.2 in another, so retrieval returns both and ranks them by wording.
Reconstruction is done by hand, every time
Asked what governed in November, someone opens four files and a signature page and works it out. The answer is correct, takes an hour, and is not written back anywhere — so the next person does it again.
Authority moves and the record does not
The signatory matrix is a spreadsheet with a last-modified date and no history. Who was empowered to sign in March is not recoverable from it, only from memory.
Mostly valid time. Then three others.
Every application built on CortexDB is a different mixture of the same five memory jobs. That mixture is what makes this a different piece of software from the one next to it, even though the infrastructure underneath is identical.
Every memory carries when it was true and when you learned it
Two clocks. An amendment does not overwrite the clause it amends — it records that the earlier text stopped governing, and from when. That is what turns “as of 18 April” from a research task into a query.
- Ask what governed on a date, not just what governs.
- Revisions keep their provenance instead of replacing it.
- Superseded text stops being returned as current.
A lookalike clause is not this clause
Standard language appears in every contract you have. Selection weighs which agreement it belongs to, not how closely it matches.
What was argued last time this came up
The position taken, the counterparty's response, and how it settled stay attached to the clause that caused it.
One record, both desks
The contracting agent and the litigation agent read the same history instead of maintaining two.
That is four of the five. The fifth — persistent memory, where an individual's preferences and corrections survive — leads for assistant and commerce agents instead. Same engine, different mixture.
Where the agreement actually lives
Rarely in one repository. The executed PDF, the amendment that changed it, the email where the extension was agreed, and the ticket where the exception was approved.
01 Wire the sources
Vendor connectors ship as extras on the connector package.
pip install 'cortexdb-connectors[slack,notion]'02 Write the amendment against what it supersedes
observed_at is the date the change took effect, not the date you indexed the file. That distinction is the entire product here.
curl -X POST http://localhost:3141/v1/experience \
-H 'Content-Type: application/json' \
-d '{
"scope": "ws:legal/contracts",
"modality": "observation",
"content": {
"kind": "text",
"text": "Amendment 3 raised the clause 7.2 liability cap from 1x to 2x annual fees, effective on execution. Supersedes the cap in the executed MSA."
},
"context": { "observed_at": "2026-08-27T09:14:00Z" },
"idempotency_key": "drive:msa-acme:amendment-3"
}'03 Recall what governed, with its provenance
Facts come back carrying the validity recorded alongside them, so the answer can name which instrument it relied on.
from cortexdb.v1 import V1Client
client = V1Client(api_url="http://localhost:3141", actor="agent:contract-desk")
pack = client.recall(
scope="ws:legal/contracts",
query=counsel_question,
view="holistic",
include=["facts", "beliefs", "events"],
budgets={"max_tokens": 1100},
)
reply = llm.respond(counsel_question, context=pack["context_block"])pip install cortexdbai
or npm i cortexdbai — both ship the same client.
Full endpoint reference in the
docs.
Privileged material does not belong in someone else's index
Contracts, positions, settlement history and the reasoning behind them. Much of it is privileged, and privilege is not a property that survives being embedded by a third-party API.
CortexDB is a single container. It runs on your own hardware, inside your VPC, or air-gapped — with the embedding and generation calls pointed at a local model if nothing may leave at all.
docker run -d --name cortexdb -p 3141:3141 -v cortexdb-data:/data cortexdb/cortexdb:latestDeployment options, model routing and the enterprise notes are on the download page.
The same infrastructure, a different mixture
Each of these leads with a different memory job, pulls from a different set of sources, and needs a different call. All of them, by job.
Bring the workflow that gets it wrong.
Not the demo. The one that answered from the wrong version, or asked a question it had already been told the answer to.