Mycelic connects evidence across enterprise systems so teams can investigate problems while sensitive records stay where they are.
On-time delivery 96% → 67%for orders packed from Mon 28 Sep
Stays here: names, addresses
Stays here: names, addresses, rates
Pickup moved to 15:00 on 28 Sep, the time the last packing wave starts. Those 375 orders miss the truck and arrive a day late.
Explains 375 of 410 late deliveries.
“The new route planner caused it.” Only link: the same start date.
To answer: connect the route planner’s log.
Sample data: a scripted walkthrough with one warehouse and one carrier, not a customer result. Click any record or result to see its evidence.
Memory keeps each system’s records where they are. Discovery puts one question to every system and connects the answers. Operations Control issues a finding only when the records support it.
In the demo: each system answered from its own records. Only order ids, dates, times and counts crossed. Names, addresses and rates stayed put.
NeuralGraph is our open-source memory for a single agent: one Rust binary, one SQLite file, local models. Sharing memory across enterprise systems is in development.
In the demo: neither system could explain the delay on its own. The question went to both, and joining their records on order id showed that all 375 last-wave orders were among the 410 late deliveries.
Our current Discovery experiment is an Open Claw society, now living through 1,000 years of history. It is still running, and there are no results yet.
In the demo: the finding was issued because four records link and agree. The route-planner claim was withheld because no record links it to a late order.
First applied in TraceNOC, below. TraceNOC is a research prototype, and its code is not public.
TraceNOC applies the Operations Control evidence check to network incidents. Memory and Discovery were not part of this study. We rebuilt the 30 August 2020 CenturyLink outage from public RIPE RIS routing updates and RIPE Atlas probe measurements, then asked agents when it started, which networks it hit and when it ended.
An answer counts only if every record it needs was observed before the question’s cutoff. Otherwise the agent should withhold and say what is missing. The evidence planner decides which records to fetch. It scans routing activity hour by hour and looks closer where activity turns elevated or calm.
Correct decisions, counting justified withholds: 86 vs 55 of 180 (p < 0.001). 680 of 720 scheduled runs completed.
What didn’t work. A swarm of agents with the same planner showed no detected accuracy advantage (75 vs 71 correct decisions, p = 0.54) and used 2–3× the tokens. With the planner, the agent also withheld correctly less often: on 17 of the 90 questions that needed it, against 36 without the planner.
Each result is labelled by the kind of evidence behind it. None of them comes from a customer deployment.
| Result | Type | What it covers | Source |
|---|---|---|---|
| TraceNOC: 21.1% → 76.7% of answerable questions correct with the evidence planner | Benchmark | The 2020 CenturyLink outage, rebuilt from public RIPE data. One incident, a 9B model, 680 of 720 runs. | CenturyLink study |
| NeuralGraph: routing each question to the memories of the person it asks about, plus question/reply pair back-fill, lifted LoCoMo accuracy from 64.9% to 73.8% | Benchmark | Public LoCoMo dataset: all 282 questions of one category, drawn from all 10 conversations. Research retrieval stack, not the installed neuralgraph-memory server. Gemma-4-e4b lenient judge; the gain held under four judges. The full end-to-end run has covered 744 of 1,540 questions so far. | Research notes |
| Discovery: an Open Claw society, currently living through 1,000 years of history | In progress | Still running. No results yet. | Research |
| Pilot outcomes | Pilot | None reported yet. | — |
Step through Operations Control, Memory and Discovery one at a time, with the research behind each.
On 30 August 2020, a backbone routing disruption hit CenturyLink (AS3356). TraceNOC replays it from public RIPE RIS and RIPE Atlas records. An agent may issue a finding only when every piece of evidence it requires is observed.
Replay of the four answerable questions from the paper’s six-question development study (local Qwen 9B). The central agent’s decisions are as logged; with the planner, all four were supported-correct. Development results are not pooled with the 9B Modal evaluation below.
The planner lifted the central agent from 21.1% to 76.7% on answerable questions (86 vs 55 correct decisions, p < 0.001). With the same planner, the swarm showed no detected accuracy advantage (75 vs 71 correct decisions, p = 0.54) and used 2–3× the tokens. The original swarm withheld correctly on every unanswerable question it completed (87/87) but answered few, so the two metrics rank the policies differently.
NeuralGraph turns each conversation turn into memories, entities and typed relations in one local SQLite file. Conversation turns from agents such as Claude Code, Codex, Claude Desktop and Cursor are the first source; more sources are planned. Sharing memory across an organization is in development.
Explore NeuralGraph Only the Memory layer is open source today. The Discovery and Operations Control code is not public.
When no single view can answer, Discovery sends the question to the agents that hold the evidence and asks again from what comes back. Our current Discovery experiment is an Open Claw society, now living through 1,000 years of history. It is still running, and there are no results yet.
Four independent parts of an organization each hold a fragment of the picture. No single one is enough to reveal the larger pattern.
Each result is labelled by the kind of evidence behind it.
Central versus swarm diagnosis under an evidence gate
We rebuilt the 2020 CenturyLink outage from public RIPE RIS and RIPE Atlas data. On 90 answerable questions, an evidence planner raised the same 9B model from 21.1% to 76.7% correct. With the same planner, a three-role swarm showed no detected accuracy advantage and used 2–3× the tokens. One historical incident; a research prototype with no public code.
Open-source (MIT) memory for Claude and other MCP agents: one Rust binary, one SQLite file, local models. On the public LoCoMo benchmark, routing each question to the memories of the person it asks about, plus question/reply pair back-fill, lifted end-to-end accuracy from 64.9% to 73.8% on 282 questions (research retrieval stack, Gemma-4-e4b lenient judge); the gain held under four judges. The 282 questions come from all 10 conversations; the full end-to-end run has covered 744 of 1,540 questions so far.
An Open Claw society, currently living through 1,000 years of history. The run is still going, and there are no results to report yet.
We share evaluation details with research collaborators and with teams considering a pilot.
Discuss research or a pilotCentral versus swarm diagnosis under an evidence gate
Loading the paper…
Name the problem and the systems that hold the evidence.