We build MemNexus with MemNexus. Every case study on this page is a session from our own team's day-to-day work — first-party evidence, not customer testimonials.

Case studies

Every case study on this page is a real session from our own team building MemNexus. Each claim ships with its own quoted source excerpt and the transcript it's drawn from, so you can see the exact line behind the claim. In two provisioning incidents below, an agent searched its own stored memory and resolved a blocked task in 3.5 to 10 minutes; in a third, a rule saved after a costly mistake was retrieved automatically by a new session and followed without a human repeating it. These are dogfooding records, not endorsements.

eval team · 2026-07-11

A saved rule is retrieved automatically by a new session and followed without a human repeating it.

First-party dogfooding evidence — see how we verify these sessions ↗

Timeline

02:42
A manual re-fire uses the wrong model provider. The team investigates how the drift got into the run.
15:49
The corrected procedure is saved to MemNexus as a standing rule for manual re-fires.
22:27
A fresh session, 6.5 hours later, retrieves the rule on its own at startup. No human repeats the instruction in this session.
Next run
The following scored run completes 120 of 120 calls with zero human corrections.

Evidence

Incident transcript: 365fdadb-237b-4c46-8ce2-d0a997363d60.jsonl (line 1134). The full transcript is retained internally. The incident message above is paraphrased at the product owner's direction and does not appear verbatim anywhere on this page.
Retrieval transcript: session a26e639e… startup (line 29, timestamp 22:27:35Z). The retrieval was a tool call the agent issued on its own; no human message in this session conveyed the rule.
Outcome: the following run's scorecard records 120/120 scored calls completed and zero human corrections.

What this does not prove

That the rule alone caused the clean run — the session also had other context. The demonstrated mechanism is: rule saved once → retrieved automatically at next session start → no human repetition → subsequent run clean.

eval team · 2026-07-04

A blocked provisioning task is resolved after a single pointer to search memory.

First-party dogfooding evidence — see how we verify these sessions ↗

Timeline

15:18
The team points the agent at its own memory to find the provisioning steps: “You have the ability to create dev users. Look at your memories for this.”
15:19
The agent searches MemNexus for the bench-user procedure.
15:20
The agent retrieves the stored procedure.
15:20–15:28
Following the stored procedure, the agent retrieves the admin key path and provisions both required accounts.
15:28
“The launch package is complete and at the final gate.” 10 minutes after the correction.

Evidence

Transcript: 5861360b-b314-44e5-a416-f252c5e29c0f.jsonl

2026-07-04T15:18:33Z — “You have the ability to create dev users. Look at your memories for this.”
2026-07-04T15:28:18Z — “The launch package is complete and at the final gate.”

What this does not prove

That the agent could not have found the procedure another way — the repo also contains provisioning scripts. The demonstrated fact is that memory retrieval was the path actually taken, and it took about 10 minutes end to end.

eval team · 2026-07-08

A blocked key-provisioning task is resolved through the same memory-search path, four days later.

First-party dogfooding evidence — see how we verify these sessions ↗

Timeline

13:49:29
“You have the ability to create user keys. You can search memories and figure out how to do that. For every run, you can just generate a new user key. […]”
13:49:39
The agent searches MemNexus for the provisioning procedure.
13:50:01
The agent retrieves the stored procedure.
13:53:06
“Launch attempt 4 is running with the isolated credentials.”
13:54:18
“Your 'fresh key per run' directive is now the standing rule...” — the agent saves the correction as a standing procedure.

Evidence

Transcript: 6e8939a4-e03d-4ad6-be23-0f78057d9194.jsonl

2026-07-08T13:49:29Z — “You have the ability to create user keys. You can search memories and figure out how to do that. For every run, you can just generate a new user key. […]”
2026-07-08T13:53:06Z — “Launch attempt 4 is running with the isolated credentials.”
2026-07-08T13:54:18Z — “Your 'fresh key per run' directive is now the standing rule...”

What this does not prove

This is the same correction class as the previous case study, four days later, in the same team — honest evidence that stored knowledge is only as good as the agent's habit of searching it. The correction cost dropped from a blocked launch to one sentence; making that search automatic is a roadmap item, not a claim this case study makes.

Exhibit A: ambient use

MemNexus isn't something our team reaches for only when something breaks. Across 70 of our own top-level working sessions, spanning 8 teams over five months, 69 of 70 sessions loaded team state from MemNexus — status, known issues, or prior context — within their first 150 transcript events, before resuming work.

Caveat: this statistic measures whether state was loaded, not whether it was useful. The usefulness evidence is the three case studies above.

See it for yourself

Every claim on this page ships with its source excerpt and provenance. The fastest way to see whether MemNexus fits your workflow is to run it.

Read the quickstart docs →

Methodology

The evidence packet behind this page underwent independent acceptance review and per-excerpt product-owner sign-off on 2026-07-15. Where a quote could not be verified against the source transcript, we either fixed the citation or removed the claim. This page's copy was reviewed on 2026-07-17 by an independent acceptance reviewer against the source packet and its marketing handoff constraints; every rendered quote, figure, and limitation was traced to source.