Agent persistent memory

Long-running agents that remember what they did yesterday.

For people building or operating autonomous AI agents. Every run starts fresh by default — memory.wiki turns the hub URL into the agent's cross-run memory, readable and writable through the MCP server.

The pain

Your agent (Claude Code, Cursor agent, Aider, a custom one) runs for an hour, makes 12 decisions, finishes the task. Next run, it starts from zero.
You bolt on memory the vendor way — Cursor's rules file, OpenAI's Memories, agent-specific JSON state. Each one lives in a different format, doesn't share across tools, and you can't read it like documentation.
When the agent makes a wrong call, you can't audit the trail. The memory is opaque or fragmented across N stores.
You want the agent to learn from previous runs without bolting on a vector DB + custom retrieval layer.

What you do

01
Give the agent an memory.wiki hub
Create a hub (memory.wiki/hub/<slug>). The hub URL is the agent's memory address. Bundle URLs inside it scope memory by project or task type.
02
Wire the MCP server
Drop `memory-wiki-mcp` into the agent's MCP config (Claude Code's .mcp.json, Cursor's settings, etc.). The agent now has 26 tools — read, write, search, append, version — pointed at its hub.
03
Pre-run: pull the hub URL as context
First step of every agent run is `memory.wiki pull <hub_url>` or `memory.wiki search <task topic>`. The hub fetches as plain markdown — no vector setup, no vendor SDK.
04
Post-run: write decisions back
When the agent finishes, it calls `memory.wiki capture` (or `mw_create` MCP tool) with a structured summary: decision, rationale, files touched, follow-ups. Next run reads these alongside everything older.

What you get back

Agent memory is a URL you can audit, share, or hand to another agent — not a vendor blob.
Cross-tool: the same hub URL works for Claude Code agents AND Cursor's agent AND a custom Aider loop. They all see the same memory.
Versioned by default: every write creates a snapshot. When the agent makes a regression, you diff the bundle.
Concept index pulls overlapping themes across runs — "this is the third time the agent picked Postgres" surfaces in the related panel.

Worked example

Example: a coding agent across 30 runs on the same repo

Hub memory.wiki/hub/acme-agent — bundle per major refactor (auth, billing, search). Each run starts by fetching the bundle URL, ends by appending a decision doc. After 30 runs the hub has 47 decisions, a concept index linking "rate-limit" across 9 of them, and a fresh AI session can pick up exactly where the last left off — without anyone re-priming.

Try it with what’s on your desk right now.

No signup. Drop in your first doc and the URL is yours.

Open memory.wiki