Coding agents

Claude Code

1. Build ampup-memory-client, the small binary the plugin runs (sign-in, hooks, MCP bridge), into ~/.local/bin, where the plugin looks first:
   cargo install --locked --path crates/memstore-client --root ~/.local
   
2. Run a server. On a laptop:
   AMPUP_MEMORY_EXTRACT_GOAL_BLIND=1 OPENROUTER_API_KEY=... ampup-memory serve --local
   
   With no server named and no sign-in, the hooks use http://127.0.0.1:8080. For a team server, sign in once: ampup-memory-client join https://memory.example.com --org acme.
3. Add the ampup-memory repository as a plugin marketplace and install the plugin (in Claude Code):
   /plugin marketplace add /path/to/ampup-memory
   /plugin install ampup-memory@ampup-memory
   
4. Check it: /ampup-memory:status in Claude Code, or ampup-memory-client doctor in a terminal.

What the hooks do

• A session per Claude Code session. SessionStart opens an AmpUp Memory session and puts its id in context. Compaction and resume continue it. Stop sends the conversation so far, at most every five minutes, without holding up the turn.
• Advice at the first prompt. The prompt hook asks get_advice for the task, fitted to its share of the context (max_chars), and adds the answer ahead of the recalled facts with an instruction to follow it unless the evidence contradicts it. The call runs in the background; the advice reaches Claude with the next hook.
• Advice at each step. After a turn with tool calls, or a tool call that usually ends a step (a commit, a push, a build or test run), the hook asks for the next moment. The note already served is re-read (no model call) and passed on only if a decisions model finds it useful now (AMPUP_MEMORY_ADVICE_USEFUL_MIN, 0.4). At most one ask per 30 seconds and 12 pieces of advice per session.
• Recall. Facts the memory held about the repo are added at session start; the first prompt’s best-matching facts come with the advice.
• Ended with its trace. SessionEnd ends the session with a trace (each prompt, each answer shortened, each tool call with shortened arguments and the first 200 characters of its result, under 20,000 characters) and Claude’s last answer as the outcome. Only its person can read the trace.
• Notes. The session is also remembered as a condensed note every five prompts, before compaction and at the end.
• Transcripts (opt-in). Where the org’s or team’s admins turn transcripts on (else the server’s AMPUP_MEMORY_TRANSCRIPTS=1), the hooks also keep each Claude Code, Codex and Gemini CLI session’s whole transcript, for its person alone: each prompt, each answer and each tool call with its input and result (whole up to 64 KB, else its first and last 32 KB), never thinking. It goes in parts as the session grows, at most 32 MB a session, and is never searched, extracted or read for the team. Forgetting a part forgets every version of it, and a part you forgot is never sent again. get_episode {"document": "transcripts/claude-code/<session>/"} lists a session’s parts.
• Starter memory. At session start the repo’s CLAUDE.md, .claude/CLAUDE.md, AGENTS.md, GEMINI.md, .claude/rules and .cursor/rules, and your Claude Code memory files for the repo, are imported into your own memory (?scope=mine), which no teammate reads. At most 40 files of up to 64 KB; a file is sent again only when it changed. AMPUP_MEMORY_STARTER=off imports nothing.
• Never waits on the memory. A slow or down server costs a session its context, never its time.

Commands

Variables

Codex, Gemini CLI, Cursor and Copilot

ampup-memory hooks install --harness codex      # or gemini, cursor, copilot; --url <server>; --dry-run first
ampup-memory hooks uninstall --harness codex

Hosts without hooks