The project memory remembers what has already been investigated, decided and measured in a project, and tells an agent before it does the same thing again.

The problem

A long sprint produces hundreds of tickets, research reports and measurements. A fresh agent knows none of them. It answers a question that a report from three weeks ago already answered, and nobody notices: the result looks plausible, it just cost time and money.

What it does

It builds an index over tickets, research, reports, concept files and optionally the project's code. It stays local in the project folder.
During a sprint it compares every task for a research, implementation or plan-check agent, and every new ticket, with this index.
If it finds earlier work that the task does not mention, the call is held. The message first says what to do, then lists the hits with file and line.
The requester reads the hit and acknowledges it in the task: either by pointing to the earlier work, or by explaining why the question is asked again anyway.
The same question can be asked by hand beforehand, with addp-hook graph query.

An example

A lead wants to task an agent with investigating the CSV export of the reports. The memory finds the finished ticket .addp/tickets/done/042-report-export.md (all names in this example are invented). The hook names the hit (by default as a hint; whoever turns the gate on for a role holds the call). The lead reads the ticket, sees that only the test environment was measured, and adds a line to the task: “neu untersucht, weil the measurement in 042 only covered the test environment”. The agent now starts, with the old measurement as its starting point instead of from zero.

What was measured

All figures come from measurements of 01.11.2026 on one machine (Apple M3 Max) and describe those measurements, not every project.

Trial on real data: A real session with 2,282 agent starts was replayed. The gate found both known cases of duplicated work in it (2 of 2). That is a small sample: it shows that real duplicated work is found, not how rarely the gate errs. It also fired on tasks that were no duplicated work and is therefore being tuned on real sprints.
Build: For a project with 2,473 files the complete build takes 4.8 seconds at the median.
Query by hand: In the same project 0.32 seconds at the median, under 0.41 seconds in 95 of 100 runs. With a larger index (2,664 files) it was 0.69 seconds at the median.
Size: On macOS (arm64) the program grows from 8.25 to 10.82 MB, about 2.6 MB.
Start time: Hook calls without memory involvement take about 3.6 to 4.4 ms longer.

What it is not

Not an acquittal: if the check could not run, the message says “Prüfung NICHT durchgeführt” (check NOT performed), never “no hits”.
Not a substitute for reading: it shows where earlier work lies, it does not judge it.
Not a cloud service: the index is created and stays in the project folder.

Where to go next

Configuration, commands, the forms of acknowledgement and help with problems are in the documentation included with purchase (file 25-project-memory.md in the folder docs/en). The gate itself is part of the AgentGate hook and can be switched off in an emergency.

Project memory