▤ Category

Agents & tooling

Posts in this category: 21

← Alle Beiträge
Agents & tooling 6 min

Twelve Ideas That Did Not Help

Twelve network ideas tied, lost or starved against plain search or their control, with the counts for each. Five had a prereg file. The failures in our measuring that we caught are listed beside them.

BrainNet · part 12
Agents & tooling 6 min

Rank Was Half the Answer Until One Lookup Fix

On 2026-11-13 the rank stage took 0.82 of a 1.56 second answer. A lookup fix cut it to 0.24 seconds cold, and a resident server to 0.093 seconds warm, with no change in the top five lists. Measured again on 2026-11-19: served rank median 0.100 seconds, cold rank 0.283 seconds.

BrainNet · part 8
Agents & tooling 5 min

A Commit Used to Cost Fifty Seconds

Every commit to our notes used to rebuild the whole network in 49 seconds and 8.40 GB. Four changes brought a warm refresh to about 16 seconds and 5.78 GB on 2026-11-06. On 2026-11-19 the warm refresh took 15.6 seconds but peaked at 7.40 GB, which fails our 6.5 GB line. An unchanged input costs 1.5 seconds.

BrainNet · part 11
Agents & tooling 3 min

Our Panel Said Enforced for Six Days

The panel said enforced for six days while nothing enforced it. The fix was a check that names unwired hooks, and the only cut with a measurement is also the only one a hook now denies.

Token reduction · part 3
Agents & tooling 3 min

A 7k-Token Prefix We Have Not Measured

A duplicated instruction file cost room in every request, not much money, because cached reads are cheap. The 7k figure is an estimate, and the measurement that would confirm it has not been run.

Token reduction · part 4
Agents & tooling 6 min

A Memory Line Agents Did Not Follow

A small learned memory predicted which files an agent would touch next at recall@5 0.599, against 0.534 for the usual rule. In a live A/B over 1,090 matched turns it changed nothing we could detect. Seven rules came out of the same week's rebuild and lock work.

BrainNet · part 7
Agents & tooling 5 min

Where BrainNet Still Fails

On our own vault, every network idea tied plain search at 23 of 29 and 17 of 30. Bridges, usage synapses, learned weights, Hebbian plasticity and a graph inside the model were all killed against a kill line set in advance.

BrainNet · part 5
Agents & tooling 5 min

What the Network Fixed

Spreading activation recovered 13.0 percent more held-out Wikipedia links on 1.18M articles, a local model found 42 extra hidden note links, and hub damping cut hub pages in paper answers from 22 slots to 9. The accuracy gain in brain-ask itself came from the reranker.

BrainNet · part 4
Agents & tooling 5 min

How Agents Use the Network

Agents must recall before they hypothesize. The vault's rules, a walk tool, skill hooks and a per-repo network turn the memory into something an agent uses every turn, including on an older laptop without Mojo.

BrainNet · part 6
Agents & tooling 5 min

How a Question Walks the Network

brain-ask finds seed notes by meaning, spreads activation along synapses with personalized PageRank, then lets a cross-encoder pick five. On the Wikipedia network we cut a question from 5.3 s to 0.69 s by firing 2,618 neurons instead of 1,175,258.

BrainNet · part 3
Agents & tooling 5 min

A Note Is a Neuron

Our memory is a vault of Markdown notes. BrainNet reads it as a graph, with notes as neurons and links as synapses, then merges notes, papers, code and docs into one network of 1,640,443 neurons.

BrainNet · part 1
Agents & tooling 7 min

Our Notes Do Not Have Enough Gravity

Spreading activation helped only inside a small, dense topical neighborhood. The vault scored 23 of 29 with the network on and 23 of 29 with it off, and a 5 percent slice of Wikipedia diluted the gain to 0.7 percent. Adding hub links to every note changed nothing.

BrainNet · part 2
Agents & tooling 3 min

The Tool Gate Must Own the First Move

A PreToolUse gate and a capability marker made the preferred discovery tool the first move, checked by a live wiring test. A later wrapper fix restored byte-for-byte passthrough for repeated payloads.

Gates that lie · part 2
Agents & tooling 8 min

The IPC Primitive Is a Latent Tensor

An OS design in which the message between agents is a latent tensor with an identity, and text is one adapter among several. The transport and tiering numbers are real, the fork gate is resolved, and the thesis experiment failed. That experiment asked whether an untrained model reads its own hidden states as reasoning; in the rerun, it did not.