Shipeasy
Bugs & Requests

Project memory

Every ticket your project ships or declines leaves facts behind — decisions, QA findings, the files a feature lives in. Coding agents search them before they start, so they build on what was decided instead of rediscovering it.

BetaOn this page · 5 min readWorks with · Every plan · per project

A coding agent that picks up a ticket usually starts cold: it re-reads the repo, re-asks questions your team already answered, and re-learns what QA caught last time. Project memory is the project's record of what its shipped tickets settled, kept in a form an agent can query in one call.

It is built from data the ops queue already stores — nothing new to capture, no notes to write:

KindComes fromWhat it remembers
shippeda ticket that reached Ready for QA or Resolvedwhat was asked, what the run found, the files it touched, the PR
declineda ticket closed Won't fixthe request and why it was turned down, so it isn't rebuilt blind
decisionan answered agent questionthe question, the option the team picked, the options it turned down
qaa QA note or an answered QA questionwhat "fixed" turned out to mean, what QA found still failing

The files come from the run records agents fill in while they work (the sources on an investigation), so a decision made on a checkout ticket is findable by anyone about to touch the checkout code.

Always consistent with the tickets

Memory is derived, never edited. A ticket's entries are a pure function of its own records and are rebuilt as a set whenever those records change:

  • Answer a question, add a QA note, or fill in a run record on a shipped ticket, and its entries are rebuilt right after.
  • Every other change — a status flip from the dashboard, the CLI, a bulk edit, a merged PR — is caught on the next read: before answering, a search rebuilds any shipped ticket whose entries are older than the ticket.
  • Reopen a ticket and its entries are dropped. It is no longer something the project shipped.

So there is no second copy that can drift, and nothing to back-fill: the first search on a project builds its memory from the tickets it already has.

How agents use it

Agents get memory in two ways.

Handed up front. When a ticket is delegated to a connected coding agent (Claude, Cursor or Copilot), the run's prompt carries the headlines of the entries most related to that ticket — matched on its title and description — with the ticket they came from and the files involved:

Project memory — what earlier shipped tickets settled that looks related.
Build on these decisions instead of re-deciding them, and read the full
entries before you design (`ops memory search --item 42` / the
`ops_memory_search` MCP tool):
- #31 decision: Retry policy: Three, exponential (`apps/web/src/checkout/retry.ts`)
- #28 qa: QA verification (fail): Retry fires twice when the first call succeeded
- #17 shipped: Checkout retries timed-out charges (`apps/web/src/checkout/retry.ts`)

Searched on demand. The shipeasy-ops-work skill searches memory before it designs a fix. Any agent or person can do the same:

# Around a ticket — its own words become the query; the ticket itself is left out
shipeasy ops memory search --item 42

# Free text
shipeasy ops memory search --q "checkout retry"

# By the files you are about to change — entries that touched them, their
# directory or a neighbouring one rank first
shipeasy ops memory search --paths apps/web/src/checkout/retry.ts,packages/billing/src

# Only the decisions
shipeasy ops memory search --item 42 --kind decision

Over MCP the same call is ops_memory_search with q, paths, item, kind and limit. Each entry comes back with its kind, a one-line title, a short body, the files, the PR, the ticket it belongs to, and a relevance score.

Token-light by design

Bodies are capped and a search returns 10 entries by default (up to 50), so a lookup costs a few hundred tokens instead of a crawl through the repo. Read a whole ticket with shipeasy ops get <item> when an entry is worth the detail.

How results are ranked

A search reduces its text to keywords (dropping stop-words and words under three letters) and scores every candidate:

  • a keyword in the title counts more than one in the files, which counts more than one in the body, plus a bonus for matching more of the keywords;
  • a queried path scores highest on the same file, then a file under the queried directory, then the same directory, then a neighbouring one;
  • decisions and QA findings are weighted above plain ship records — they are what an agent most often gets wrong by not knowing;
  • older entries count for less: a year-old fact counts half as much as today's.

With no query at all, the newest entries come back.

Configure it

Open Settings → Memory. Every control saves as soon as you change it.

SettingDefaultWhat it does
Keep a project memoryOnOff keeps nothing: stored entries are dropped, searches answer empty, and runs are handed no memory. On rebuilds on the next read.
Headlines per agent run4How many related entries a fired agent run is handed per ticket, from 0 to 6. At 0 agents can still search on their own.
Kinds servedAllUntick a kind to stop serving it to searches and run prompts — for example, keep declined requests out of agent prompts.
Rebuild—Drops every entry and rebuilds from the tickets now. The page shows how many entries and tickets are stored, and when it last built.

Memory stays inside its project: a search only ever reads the project the credential belongs to, and a restricted ops key can read it but never write it — there is no write route at all.

Make memory better

Memory is only as good as what runs and people leave on tickets:

  • Agents: fill in your run record as you work (ops investigations update) — a summary, findings, and the sources you read or changed. Those files are what makes a ticket findable by path.
  • Ask, don't assume. A structured question (ops questions ask) that the team answers becomes a decision every later run sees.
  • QA: a narrowing, clarification or correction note records what "fixed" means for the next person who touches that code.
Was this page helpful?
Updated November 20, 2026✎ Edit this page

On this page