A79-ai/ampup-memory

The memory for agents that gets better with use: it learns advice from the runs that went well and serves it
before the next one.
Docs · Website ·
Quickstart
A conversation is a session. Each task in it asks for advice (get_advice) and gets one skill for that kind of
task: a note learned from earlier rewarded runs, or, before any exist, a procedure compiled from your documents.
When the session ends, its transcript and outcome are judged or rewarded, and a rewarded session teaches the next
one. Underneath is a cited memory: notes, documents and conversations become searchable passages and facts, each
fact quoting its source and visible only to people who may read every source it cites. Everything is stored as
Apache Iceberg tables in a catalog and bucket you run.
## Install
AmpUp also runs it for you: hosted AmpUp Memory needs no install beyond the
plugin. To run it yourself:
bash
curl -fsSL https://raw.githubusercontent.com/A79-ai/ampup-memory/main/install.sh | sh
That puts ampup-memory (server and CLI) and ampup-memory-client (hooks for coding agents) in ~/.local/bin, from
the newest release, for macOS on Apple Silicon or Linux on
arm64/x86_64. It needs only curl (or wget) and tar; AMPUP_MEMORY_INSTALL=client installs the client alone.
The command was memstore before the rename, and memstore still works as a link to it.
To build from source instead, you need rustup (the toolchain is pinned in rust-toolchain.toml) and protoc on
your PATH:
bash
cargo install --locked --path . # ampup-memory: server and CLI
cargo install --locked --path crates/memstore-client --root ~/.local # ampup-memory-client: hooks for coding agents
For Claude Code, add this repository as a plugin marketplace and install the plugin; it downloads its client from the
releases if none is installed (coding agents):
text
/plugin marketplace add A79-ai/ampup-memory
/plugin install ampup-memory@ampup-memory
## Run it locally
bash
OPENROUTER_API_KEY=... AMPUP_MEMORY_EXTRACT_GOAL_BLIND=1 ampup-memory serve --local # ~/.ampup-memory, http://127.0.0.1:8080
claude mcp add --transport http ampup-memory http://localhost:8080/mcp
--local needs no Docker, catalog service or sign-in, and binds to 127.0.0.1. Without a model key it stores and
searches but learns nothing; any OpenAI-compatible, Anthropic, Bedrock or Vertex AI endpoint, or a local Ollama,
works instead of OpenRouter (configuration).
### Offline, on your laptop
No key and no network: local embeddings (the default) and a local model through Ollama for
everything the memory learns.
bash
ollama pull qwen2.5:3b # the default local model
ampup-memory embed warmup # fetch the embedding model (~130 MB) while you're online
AMPUP_MEMORY_LLM=ollama AMPUP_MEMORY_CLIENT_UPDATE=off ampup-memory serve --local
AMPUP_MEMORY_LLM_MODEL picks another model, AMPUP_MEMORY_LLM_URL another server (any OpenAI-compatible one:
LM Studio, llama.cpp, vLLM). Started in a terminal with nothing configured, serve --local asks once whether to
use a key, the local Ollama, or no model, and keeps the answer in ~/.ampup-memory/llm.json. Without the embedding