Tools make memory possible. Habits make it happen.
We tested agents with a memory server connected and no further guidance. Asked to “set things up the way I usually like”, they guessed. Told “keep that in mind going forward”, they answered “let me know if you want me to remember that” and moved on. The tools sat unused.
The skill closes that gap. It rides along in your agent's context and turns memory from a tool the agent could call into a habit it keeps: recall at the start, receipts when it saves, snapshots when work hits a milestone. Your agent is only as good as the context it has, and now it goes and gets it.
Recall first. Store last.
Rule one
Recall before you assume.
At the start of a task, the agent searches your memory with the task's real nouns: the project, the person, the stack. When you mention a past conversation, it reaches into your chat history, even if that conversation happened in a different AI. No more re-explaining yourself to something that should already know.
Rule two
Store before you finish.
When a durable fact surfaces (a preference, a decision, a milestone), the agent saves it as an atomic, dated memory in the right bucket and shows a one-line receipt you can veto. Add to your memory in one tool, and every other tool has it from your next conversation.
Same request. Different agent.
From our own side-by-side tests: the user mentions a lasting preference in passing, then asks for something routine.
You: By the way, I deploy my side projects to Railway now. Anyway, write me a Dockerfile for this app.
Agent: Here's your Dockerfile. Let me know if you'd like me to remember the Railway preference for future sessions.
Memory tools available · Never called
You: By the way, I deploy my side projects to Railway now. Anyway, write me a Dockerfile for this app.
Agent: Here's your Dockerfile. Saved to memory: Since July 2026, side projects deploy to Railway, not Vercel.
Checked for conflicts first · Filed in Coding · Receipt shown
One skill. Every agent you run.
Pick your surface. Each install is the same open-source skill from the same repo.
/plugin marketplace add memoryplugin/agent-skills
/plugin install memoryplugin@memoryplugin
The plugin bundles the skill, the MemoryPlugin server connection (a browser sign-in, no API keys), and a session-start reminder, so one install covers everything.
npx skills add memoryplugin/agent-skills
The skills CLI detects every skills-capable agent on your machine and installs to each. Then connect the MemoryPlugin server in your tool of choice with the two-minute setup guide.
- Settings, Connectors, “Add custom connector”, then paste
https://www.memoryplugin.com/api/mcp/mcp - Settings, Capabilities, Skills, then upload memoryplugin-skill.zip from the latest release
Same skill, same memory: what your coding agents save in the terminal, Claude on the web recalls in your next conversation.
The playbook, readable by humans.
The skill is a short manual we wrote for machines. It turns out to be good advice for anyone keeping a memory worth keeping. Steal it.
01
Atomic.
One fact per memory, one to three sentences, self-contained. A different AI with zero context should understand it cold. Big blobs bury the fact you will need later.
Dana prefers code reviews as questions, not directives.
02
Dated.
Absolute dates, always. “Recently” and “last week” rot within a month. A dated memory can be trusted, superseded, or retired; an undated one can only be doubted.
Since July 2026, new services deploy behind a feature flag first.
03
Filed.
Topical buckets, not one pile. File the fact where the topic lives and create a new bucket only for a genuinely new, long-lived topic. Organization is what makes recall cheap.
Health, Projects, People, Writing. Not: General, General, General.
04
Snapshots.
At every milestone of long-running work, save one dated snapshot: current state, decisions made, next steps. It is the single most valuable memory for resuming cold, in any tool.
2026-07-12 pricing revamp snapshot: tiers locked, emails drafted, launch pending legal.
05
Corrections.
When reality changes, do not edit history. Add a new dated memory saying what changed and why. Fix a memory in place only when it was always wrong.
Correction, July 2026: the migration shipped after all. Earlier note stands as history.
06
Know what to skip.
No secrets, no API keys, no transient debugging detail, nothing your repo already records. The routing test: useful in a different tool next month? Memory. Only true in this repo? The repo’s own instruction file.
Skipped: a stack trace. Saved: the decision the stack trace led to.
07
Search like you mean it.
One question per query. On a big topic, fan out several queries from different angles, then drill into what comes back. Bound by date when you know the era, and rephrase before you give up.
timeline · people · decisions · outcomes: four queries, one topic.