close-session

Memory Kit

Built-in memory decides what to remember. This one asks.

Your agent proposes, you say yes, it writes a dated line into plain files in your own folder. One memory per client, readable by every agent you run, nothing remembered without your yes.

"I wake up already knowing where we left off." — the agent this kit builds.

Install

Three commands in Claude Code, inside the folder where you work. Any folder will do; git is optional, and only needed if you want the wiki and rules to carry their history.

/plugin marketplace add awrshift/agent-memory-kit
/plugin install memory-kit@memory-kit
/memory-kit:setup

That is the technical part, and it happens once. From then on you only talk: work as usual, and say /memory-kit:close-session when you are done for the day. No database, no service, no extra cost. Setup reads what your folder already has and proposes before writing anything; your CLAUDE.md stays yours. Other agents install from the same repository, see Works with your agents.

claude plugin marketplace update memory-kit     # refresh the catalog
claude plugin update memory-kit@memory-kit      # the bare name resolves to nothing

Restart the session to apply. Upgrades never touch your files. Memory, handoffs and knowledge live in your folder, not in the plugin.

The problem

Every session starts from zero: yesterday you locked the brand voice for a client, today you explain it again. Your tools do remember things now, but each one keeps its own notes, outside your folder, and decides by itself what goes in. And memory nobody audits rots: a status note that froze three weeks ago still reads like today's truth, and the agent confidently acts on it.

Memory Kit's answer: one set of plain text files in your folder, written by the agent only after you agree, with a date on every line so a stale fact looks stale.

A day with it

1. Open. The agent wakes up already knowing: your hot cache, the note the last session left, and whether memory is healthy. You just continue.
2. Work. When something worth keeping comes up, the agent saves it as a dated one-liner and says "saved". Before the context is compressed, it has to save state. Weakening an existing test needs your yes, a force push or a hard reset is stopped before it runs, and your .env and keys stay out of the transcript.
3. Close. Say you are done for the day. The close ritual audits instead of dumping logs: "you rejected em-dashes on four different dates, make it a rule?" You say yes, it writes, and leaves the note tomorrow's session opens with.

No slash commands in any of this. You talk, the skills trigger from what you said:

The whole session, uncut except for the model's thinking time: 68-second video. Seeded folder, real Claude Code, reproducible from tools/demo/.

Why not the built-in memory?

Claude Code, Cursor and Copilot all ship memory now, on by default. It is effortless, and it is the vendor's silo. The kit is the opposite trade.

Running both means two writers and two truths, so /memory-kit:setup asks you to pick. Either answer is legitimate.

This is the moment the table is about. "That's all for today, wrap up" — the close ritual finds the pattern on four dates and asks; the rule exists only after the yes:

And the next morning, a new session, a plain question:

Many clients, one agent

The line is memory vs paperwork. What the agent learned is shared across everything you do: patterns, knowledge, rules. What the work produced belongs to one client: projects/<client>/ holds that client's backlog, specs, research, decisions and QA records. Say "we're working on Nestlé" and the agent loads that scope only. One folder per client, or one repository per client, both work.

A pattern's journey: noticed, saved as a dated line, repeats on 3+ dates, the agent proposes promotion, your "yes" makes it a knowledge article or a rule, and the raw lines are pruned. Observation, candidate, law. You approve every step.

"Now IKEA" — the agent reads that client's folder, and a schedule change becomes a dated line without leaving the conversation:

Works with your agents

codex plugin marketplace add awrshift/agent-memory-kit && codex plugin add memory-kit@memory-kit
cursor-agent plugin marketplace add https://github.com/awrshift/agent-memory-kit
copilot plugin marketplace add awrshift/agent-memory-kit && copilot plugin install memory-kit@memory-kit

OpenCode takes one line in opencode.json: "plugin": ["memory-kit@git+https://github.com/awrshift/agent-memory-kit.git"]

Hosts that cannot run hooks get the same discipline as an always-loaded instruction: /memory-kit:setup offers a small protocol block for your AGENTS.md. Every claim is probed, dated and labeled in docs/specs/, degradation stated, never hidden.

Private by default

Everything is plain text on your machine; nothing leaves. The hot cache and the session notes are gitignored by default, so they stay