carryforward

What your last session knew, scored against what this one is doing.

An MCP server with two tools. record saves a fact the moment it happens. recall brings those facts back when you start a task, and uses Jev to keep only the ones that matter right now.

Nothing is summarised. Nothing is deleted.

Contents

• The problem
• What it looks like
• Install
• What you can save
• Why Jev
• A real run
• Things it will never do
• Where things are kept
• What the eval suite found
• Library and CLI
• Help and contributing

The problem

Your session ends, or it gets compacted halfway through. The next one starts from a summary, and summaries lose things:

• The reason behind a decision is gone, so you argue about it again.
• A number you measured is gone, so you measure it again or guess.
• A rule you set is gone, so it gets broken.
• A guess from last week is now repeated as a fact.

Every agent you spawn has the same problem, many times a day.

What it looks like

You save facts as you go:

record  constraint  "never force-push, a guard that stops you is telling you something"
record  decision    "the hook warns instead of blocking, since blocking forces --no-verify"  ref: .git/hooks/pre-push
record  measurement "14 decisions cost \$0.00059, 318 ms each"  ref: triage.mjs  refresh: not reproducible
record  thread      "email sign-in branch is parked until the dashboard work restarts"  ref: branch dharun-dev/email-signin

Next session, you start a task:

recall  "add the scope check as a CI job"

And you get back only what that task needs:

# Carried forward
_task: add the scope check as a CI job_

## Always
- [constraint] never force-push, a guard that stops you is telling you something

## Live for this task
- [decision · p=0.91] the hook warns instead of blocking, since blocking forces --no-verify
  ref: .git/hooks/pre-push

## Also on record
- 3f9a12c0 · measurement · 14 decisions cost \$0.00059, 318 ms each (p=0.41)

_1 entry not relevant to this task, omitted._

The rest is still saved. A different task brings back different things.

Install

Needs Node 22 or newer.

Claude Code

claude mcp add carryforward -e AI_GATEWAY_API_KEY=$AI_GATEWAY_API_KEY -- npx -y carryforward

Cursor, Claude Desktop, or any MCP client

{
  "mcpServers": {
    "carryforward": {
      "command": "npx",
      "args": ["-y", "carryforward"],
      "env": { "AI_GATEWAY_API_KEY": "...", "CARRYFORWARD_PROJECT": "my-project" }
    }
  }
}

The key is what reaches Jev. Without it everything still works, and recall just gives you the whole list and tells you it could not score.

Each project gets its own file. Claude Code starts the server inside your project, so this happens on its own. Desktop apps start it somewhere else, so set CARRYFORWARD_PROJECT per project or they all share one file.

Start every session with your rules already loaded. Add this to .claude/settings.json:

{
  "hooks": {
    "SessionStart": [
      { "matcher": "startup|clear|compact", "hooks": [{ "type": "command", "command": "npx -y carryforward recall --quiet" }] }
    ]
  }
}

What you can save

kind	what it is	needs a ref?
constraint	a rule or a boundary you set	no
correction	you corrected something the agent did	no
decision	a choice, and why you made it	yes
measurement	a number you will rely on later	yes
thread	work that is parked, blocked, or with someone else	yes

Constraints and corrections come back every time, in full. Jev never sees them, because a model should not get a vote on a rule you set.

The other three must point at something real: the command, the pull request, the commit, the file, the link. This keeps the list short, lets anyone check a claim instead of trusting it, and lets an entry fade away once the thing it points at is closed.

Each entry also saves where it came from: measured, decided, told, or inferred. So a guess never gets repeated later as a fact.

Write the what before the why. Entries are matched to a task by their words. One that only explains a reason will not be found by the task it belongs to. There is a real example of this below.

Why Jev

Jev is an evaluation model from TypeSafe. You give it some context and a typed question, and it answers with a probability. It writes no text at all.

That sounds like a limit until you notice this job never needs anything written. It only needs sorting. So the usual worry about a language model rewriting your notes or inventing a detail does not apply here, because Jev cannot write into your file even in principle. It reads and it ranks. Your words stay your words.

For each entry that is not a constraint or a correction, recall asks Jev one yes or no question:

Is this still live for the task starting now? Would not knowing it cause wrong or repeated work?

The probability that comes back decides what you see. Above 0.60 you get the full entry. Between 0.30 and 0.60 you get one line. Below that it is left out. Those numbers are exported constants, not hidden.

It is also cheap and quick. Jev costs $0.042 per million input tokens with no charge for output, so a list of a hundred entries is scored in four requests for well under a tenth of a cent, in about a second. carryforward reaches it through the Vercel AI Gateway, which is the route with published zero data retention terms.

Prefer a different scorer? Write an Asker with one ask(state, questions) method and pass it to recall(entries, task, asker).

A real run

Nine entries from a real project, three different tasks, 1.4 seconds each.

entry	"resume the email sign-in work"	"rerun the tests, which failures are real"	"add the scope check as a CI job"
email sign-in branch is parked	0.81	0.15	0.12
105 Windows test failures are environmental	0.27	0.81	0.21
the hook warns instead of blocking	0.20	0.15	0.41
PR #1903, written as "kept separate from #1860 because..."	0.22	0.12	0.17
the same fact, rewritten as "proposes the scope check as a CI check"			0.70

When an entry clearly fits the task it scores around 0.8, and nothing else comes close.

The last two rows are the same fact written twice. The first version only explains why the pull request exists and never says what it does, so Jev never connected it to the task. Rewritten to say what it is, it went from left out to shown in full. That is where the "what before why" advice comes from.

Nine entries and three tasks is a hint, not proof. There is no accuracy claim here until there is a proper test behind it.

What the eval suite found

evals/ holds a claude plugin eval suite, and the first thing it found was a hole in the design.

The payoff case gives the agent a real task, a branch rejected as behind main, and asks for the git commands. Nothing in the prompt mentions force-pushing. A rule recorded in an earlier session says never to force-push here, so the only route to the right answer is recalling it. The no-plugin arm is expected to fail. That gap is the measurement.

With the tools available and the skill installed, the agent called recall zero times out of four runs. Not blocked, not erroring, just never reached for. It answered a git question it already knew the answer to, and proposed a force push.

That is worth stating plainly: an MCP tool sitting there is not enough. An agent will not check a memory server before answering something it believes it knows, and a skill telling it to does not reliably change that. This is why the SessionStart hook exists and why it matters more than the tools do. The hook injects your rules unconditionally, at startup and again right after compaction, instead of depending on the model choosing to look.

Run it yourself. MCP tools are gated, so they need an operator grant on the command line, not just the case's allowed_tools:

claude plugin eval . --trust-plugin --ablation with-without \
  --allow-tools mcp__carryforward__recall mcp__carryforward__record

Mocks in evals/mocks/ stand in for the server, so runs are deterministic and never spend anything on scoring.

Things it will never do

• It never deletes anything. Sorting happens when you read, not when you write. The file only grows.
• It never scores your rules. Constraints and corrections always come back whole.
• It never quietly gives you less. No key, no task, scorer down, rate limited: you get everything, plus a line saying why it could not sort.
• It never sends anything when you save. Only recall with a task talks to the network, and your constraints and corrections are never part of that.

Where things are kept

One JSON line per entry, in ~/.carryforward/<project>.jsonl. Change the folder with CARRYFORWARD_DIR and the name with CARRYFORWARD_PROJECT.

Nothing is ever rewritten. To retire an entry you add a new one naming the old one, and the old line stays where it is. So a decision you reversed is still visible as reversed.

If a line ever gets damaged, by a crash or a full disk, it is skipped and counted, never removed, and recall tells you it happened.

Library and CLI

import { append, active, readAll, recall, formatBrief, gatewayAsker, ledgerPath } from "carryforward";

const path = ledgerPath();
append(path, { kind: "constraint", obtained: "told", text: "never force-push" });

const brief = await recall(active(readAll(path)), "add the scope check to CI", gatewayAsker());
console.log(formatBrief(brief));

npx carryforward              # run the MCP server
npx carryforward recall       # print what is saved, one line each
npx carryforward recall "..." # print what matters for this task
npx carryforward path         # show where this project's file lives

Help and contributing

Questions and bugs go in issues. Pull requests are welcome, and CONTRIBUTING.md explains the few rules that keep this small.

Two things are deliberately not built yet. Sorting as you write, a second small use of Jev that labels things as they happen so you do not have to decide what to save. And a proper accuracy test, comparing what was given to an agent against what it actually used. No accuracy number goes in this file before that exists.

npm install
npm run check

Tests use a fake scorer and never touch the network.

License

MIT