Project Brain

  1. › opening Project Brain
  2. › synthetic demo
  3. › loading workspace
Skip to content
Project Brainmemory for your AI agents

Workspace

DashboardProjectsMemoryQualityActivitySettingsSetup

Read-only example → install your own

Demo dataNumbers are examples until the API is connected.

Help

Workspace/Dashboard

Demo data · synthetic examples

Read-only preview. No repository or model is connected; the numbers and project knowledge are examples.

Install your own →

Loading this screen…

Dashboard

Whether your agents get current project memory, and what to fix next.

Loading measurements…

4 projects: 2 behind their codebilling-service — 41 changes not read yet · project-brain — 3 changes not read yet

Open index runs for billing-service

Memory map · billing-service

How much of your code Brain remembers

67.4%

Each square represents indexed code. Point at a square or select a module to inspect its coverage.

  • remembered13,370
  • changed since3,900
  • not read yet2,544
  • skipped on purpose606

Counted in . Squares are proportional, one square ≈ 24 pieces.

Brain is reading the latest changes right now

Project memory

How much of each project's code Brain can hand to agents right now.

Projects
ProjectStatusCode agents can use now
project-brainmaster · d57f0143 changes not read yet
18,263/18,903
640 outdated
docs-sitemain · b113acaUp to date
4,150/4,150
All indexed code is current
mobile-appmaster · bec759dUp to date
12,340/12,340
All indexed code is current
billing-servicemain · 9f949e941 changes not read yet
5,567/9,687
4,120 outdated

Value to agents

Who uses Brain and whether the memory you recorded reaches them, last 7 days

Context packs
Memory reach
79.5%
31 of 39 packs carried a rule, decision or learning
Packs delivered
42
37 attributed to an agent · 5 without a known caller
Typical pack
6,120 tok
9 files at the median · 15,990 tok in the largest 5%
Rules reaching agents
4 of 4
12 reviews cited a rule · 0 reached nobody
AgentRequestsPacksPack rateWith memoryLast seen
claude-codemcp · http · 64 failed6,204196.6%1602 Oct 2026, 13:58 UTC
codexmcp · 51 failed4,981148.1%1002 Oct 2026, 12:58 UTC
cursormcp · 48 failed3,10266.1%402 Oct 2026, 12:31 UTC
devinmcp · 12 failed44335.7%102 Oct 2026, 10:55 UTC

Agent requests

Search and context calls from agents, last 7 days

24h7d30d
Requests
14,730
from 4 agents · 88 without a known agent
Context delivered
97.9%
598 of 611 context pack requests answered
Failed
61
0.4% of requests, incl. timeouts
Slowest searches
184 ms
95% of searches were faster · context packs 2,310 ms
2,105
25 Sept 2026, 14:02 UTC01 Oct 2026, 14:02 UTC
View exact measurements
Intervals in UTC. A dash means not measured.
IntervalRequestsFailedp95, ms
25 Sept 2026, 14:02 UTC2,1059168
26 Sept 2026, 14:02 UTC2,1059205
27 Sept 2026, 14:02 UTC2,1049182
28 Sept 2026, 14:02 UTC2,1049219
29 Sept 2026, 14:02 UTC2,1049196
30 Sept 2026, 14:02 UTC2,1048173
01 Oct 2026, 14:02 UTC2,1048210

Requests Failures Every request

Background jobs

Code reads and checks Brain ran recently.

Activity
  • 14:02:11Reading recent changesproject-brain · 14 / 31 filesrunning
  • 14:02:11Making new code searchableproject-brainqueued
  • 13:58:40Building a context packdocs-site · took 4.2scompleted

Search quality

How often search finds the right file, from the last saved test run.

Quality
Demo data96 test questions with known right answers · 45 test tasks
Right file in top 587.4%Questions where a right file was among the first 5 results
Ranking score0.704Mean reciprocal rank: higher is better; 1.000 means a right file came first for every question

Measured 02 Oct 2026, 03:00 UTC. A past test run, not live traffic; the period selector does not change it.

How these numbers are measured
  • Agent requests are recorded HTTP /search, /context and MCP search_code, prepare_task_context calls, failures and timeouts included. Collection began 24 Sept 2026, 09:12 UTC. A dash means not measured, never zero.
  • Memory reach is the share of context packs that carried at least one rule, decision or learning. Only packs whose contents Brain recorded are counted; recording began 24 Sept 2026, 09:12 UTC. Older packs count as not measured, never as empty.
  • Pack rate is how many durable packs an agent got per context call. It is not a count of its tasks: an agent can take context without a pack being stored, which is a normal outcome.
  • Rules reaching agents counts rules that went into at least one pack. “Cited a rule” counts diff reviews that caught something against one, recorded since 24 Sept 2026, 09:12 UTC. A review that failed to run records nothing, so it never reads as a clean diff.
  • Context delivered is technical success: the call returned context. It does not say whether the context was right.
  • Response time is server handling time, failures included; authentication and transport are excluded.
  • Code agents can use now counts indexed pieces of code that match the latest revision. It comes from the latest index and does not depend on the selected period.
  • Slowest searches is the 95th percentile: 95% of searches finished faster than this.
  • Search quality is a saved benchmark on a fixed question set. Task usefulness and token, time or cost savings are not measured.
  • A failed background job is a record of what happened, not proof of an open incident.

Observed 02 Oct 2026, 14:02 UTC.