keep
Open-source memory and knowledge for AI agents, written in Rust
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Check it deeper on your machine
It has changed since its scan on 21 Nov 2026.
As a Claude Code plugin, install it and then ask Claude Code:
/plugin marketplace add keithmackay/mackayi/plugin install scanskillsafety@mackayiDo a deep scan of https://github.com/TulipFarm/keep with scanskillsafety
Or from a terminal, clone it and run it:
git clone https://github.com/keithmackay/scanskillsafety-skill ~/.claude/skills/scanskillsafetynode ~/.claude/skills/scanskillsafety/cli.cjs --deep 'https://github.com/TulipFarm/keep'
It checks the exact commit: every file, its packages against OSV (api.osv.dev), how it compares with this page's scan (findsafeskills.com receives the repository's address), and, if you agree, Claude reading every file with your own Claude subscription. It never runs the code.
Badge for your README
Markdown
[](/listings/tulipfarm--keep)HTML
<a href="/listings/tulipfarm--keep"><img src="/api/listings/tulipfarm--keep/badge.svg" alt="findsafeskills safety rating" /></a>Install
Install instructions from the README
As written by the repo's author. findsafeskills has not verified these steps.
These steps run a script from the network that findsafeskills does not scan.
curl -fsSL https://keep.tulipfarm.ai/install.sh | sh
The script asks which local database to use (in memory, Postgres + pgvector, or Postgres + Qdrant), builds Keep and connects Claude Code, Codex and Copilot CLI. Prefer to let your agent do it? Tell it: `Install https://keep.tulipfarm.ai/install.md`. For a container, use `ghcr.io/tulipfarm/keep`. See [Connecting coding agents](#connecting-coding-agents) for the manual steps.
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
README.md · README · 189 lines · 1 finding marked
- <p align="center">
- <a href="https://keep.tulipfarm.ai"><img src="docs/assets/banner.svg" alt="Keep: memory your agents can't quietly rewrite" width="100%"></a>
- </p>
- <p align="center">
- <a href="https://github.com/TulipFarm/keep/actions/workflows/ci.yml"><img alt="CI" src="https://img.shields.io/github/actions/workflow/status/TulipFarm/keep/ci.yml?branch=main&label=CI&style=flat-square"></a>
- <a href="https://github.com/TulipFarm/keep/actions/workflows/image.yml"><img alt="Image" src="https://img.shields.io/github/actions/workflow/status/TulipFarm/keep/image.yml?branch=main&label=image&style=flat-square"></a>
- <a href="LICENSE"><img alt="License: Apache-2.0" src="https://img.shields.io/badge/license-Apache--2.0-0a0a0a?style=flat-square"></a>
- <img alt="Written in Rust" src="https://img.shields.io/badge/written%20in-Rust-dea584?style=flat-square&logo=rust&logoColor=white">
- <img alt="MCP server" src="https://img.shields.io/badge/MCP-server-0a0a0a?style=flat-square">
- <img alt="Status: research preview" src="https://img.shields.io/badge/status-research%20preview-e5311d?style=flat-square">
- <a href="https://github.com/TulipFarm/keep/stargazers"><img alt="GitHub stars" src="https://img.shields.io/github/stars/TulipFarm/keep?style=flat-square&logo=github"></a>
- </p>
- <p align="center">
- <a href="https://keep.tulipfarm.ai"><b>Website</b></a> ·
- <a href="docs/server.md">Docs</a> ·
- <a href="https://github.com/TulipFarm/keep">GitHub</a> ·
- <a href="https://github.com/TulipFarm/keep/issues">Issues</a> ·
- <a href="https://tulipfarm.ai">TulipFarm</a>
- </p>
- # Keep
- Keep is an open-source memory and knowledge system for AI agents, written in
- Rust. It is built around two primitives:
- - **Memory files** — small, curated Markdown files an agent reads at the start
- of every session (preferences, standing instructions, project conventions).
- Every change is a revision, so memory can be reviewed, diffed and rolled back.
- - **Knowledge** — an append-only record of what happened (episodes) and the
- facts derived from it, searchable on demand. Facts keep their provenance and
- validity over time instead of being silently overwritten.
- ## Quick start
- ```sh
- curl -fsSL https://keep.tulipfarm.ai/install.sh | sh
- Note: Pipes a downloaded script directly into a shell (curl | sh / wget | bash)
- Line 38 on GitHub
- ```
- The script asks which local database to use (in memory, Postgres + pgvector,
- or Postgres + Qdrant), builds Keep and connects Claude Code, Codex and Copilot
- CLI. Prefer to let your agent do it? Tell it:
- `Install https://keep.tulipfarm.ai/install.md`. For a container, use
- `ghcr.io/tulipfarm/keep`. See [Connecting coding agents](#connecting-coding-agents)
- for the manual steps.
- ## Status: research first
- Keep is at the very start. Before building the engine, we are building the
- benchmark harness that will decide whether it is worth building: a
- preregistered, reproducible way to compare Keep against simple baselines
- (full context, BM25, dense retrieval, plain notes) on public long-term memory
- benchmarks. If Keep does not beat those baselines, we will say so.
- What exists today:
- | Crate | Purpose |
- | --- | --- |
- | `keep-core` | Domain types for Memory files, revisions, episodes and facts; content hashing |
- | `keep-providers` | Chat, embedding and tokenizer traits; deterministic mock providers; an OpenAI-compatible client that refuses to dispatch unless explicitly allowed; budget caps |
- | `keep-store` | Lexical (BM25) and vector index contracts with in-memory implementations |
- | `keep-postgres` | pgvector adapter |
- | `keep-qdrant` | Qdrant adapter |
- | `keep-memory` | The Memory primitive: lossless Markdown entries, revisions, compare-and-swap edits, trash, undo, forget, context assembly, export and import |
- | `keep-ledger` | Postgres storage for Memory, with its own migrations |
- | `keep-knowledge` | The Knowledge primitive: bi-temporal facts with evidence, hybrid search, a permission-safe book index and forgetting. See [docs/knowledge.md](docs/knowledge.md) |
- | `keep-bench` | The benchmark harness library and `keep-bench` CLI |
- | `keep-server` | The `keep` binary: Memory and Knowledge for coding agents over MCP (stdio) and HTTP. See [docs/server.md](docs/server.md) |
- Every Keep ablation arm runs in the harness: Memory only (`keep-m*`),
- Knowledge only (K0, K1) and both together (Mk0, Mk1). See
- [docs/memory.md](docs/memory.md) for how Memory works.
- ## Running the harness
- Requires Rust (the pinned toolchain in `rust-toolchain.toml` installs
- automatically through rustup) and, optionally, Docker.
- ```sh
- cargo test --workspace # unit and integration tests, offline
- cargo run -p keep-bench -- dry-run fixtures/manifests/longmemeval-mock.toml
- cargo run -p keep-bench -- run fixtures/manifests/longmemeval-mock.toml
- cargo run -p keep-bench -- run fixtures/manifests/longmemeval-memory-mock.toml
- cargo run -p keep-bench -- run fixtures/manifests/ablation-mock.toml
- cargo run -p keep-bench -- report results/<run-id>
- ```
- Everything above runs offline with mock models and synthetic fixtures. Mock
- output measures the plumbing, not memory quality: never quote it as a result.
- Real benchmark data is downloaded separately and never committed:
- ```sh
- scripts/fetch-datasets.sh # LoCoMo, LongMemEval-S, oracle → data/
- cargo run --release -p keep-bench -- dry-run manifests/longmemeval-s-dev.toml
- ```
- The manifests in `manifests/` are templates: model pins are placeholders and
- budget caps are zero, so they cannot call a model until someone fills them in
- deliberately. Remote model calls need `dispatch.allow_model_calls = true` in
- the manifest **and** `--allow-model-calls` on the command line.
- For development without API keys, `manifests/dev/` runs answer and judge
- calls through a locally installed, signed-in GitHub Copilot CLI and embeds
- with a pinned local ONNX model:
- ```sh
- cargo run --release -p keep-bench --features local-embeddings -- \
- dry-run manifests/dev/locomo-copilot-dev.toml
- cargo run --release -p keep-bench --features local-embeddings -- \
- run manifests/dev/locomo-copilot-dev.toml --allow-model-calls
- ```
- These runs are **non-canonical**: the CLI adds its own hidden system prompt
- and does not pin the upstream model. Every artifact is stamped as such, the
- run stops at the manifest's AI-credit cap, and `keep-bench compare` refuses
- them unless `--allow-non-canonical` asks for an internal comparison, whose
- every output is stamped INTERNAL / NON-CANONICAL. See the methodology for
- details.
- The canonical G0 manifests in `manifests/g0/` pin gpt-4o-mini (answers) and
- gpt-4o (judge) served by Azure OpenAI. `scripts/run-g0.sh` checks the
- credentials file, prints the dry-run estimate and runs them after
- confirmation; see [docs/benchmarks/running-g0.md](docs/benchmarks/running-g0.md).
- `scripts/run-g0-copilot.sh` defaults to the v5 Copilot CLI experiments
- (`manifests/g0/*-copilot-v5.toml`) as an internal, non-canonical check.
- Set `KEEP_G0_VERSION=1`, `2`, `3` or `4` to reproduce an earlier version,
- `6` for the v6 experiments (answer-v6 working notes and consistent
- judge verdicts), or `7` for the v7 experiments (answer-v7 world-knowledge
- bridging and specific wording).
- G2 is the gate-eligible run. `manifests/g2/` compares keep-mk7 with
- full-context-v7 and dense-6k-v7 (keep-mk6 is reported, not gated), with
- three seeds on both storage profiles against one pinned Azure `gpt-6-luna`
- deployment, and has separate sealed-holdout confirmation manifests. The
- `-v2` manifests (`scripts/run-g2-azure.sh --set v2`) swap the candidate
- for keep-mk8 and report keep-mk7.
- `scripts/run-g2-azure.sh` requires a dry-run estimate and a `--max-usd` cap;
- see [docs/benchmarks/running-g2.md](docs/benchmarks/running-g2.md).
- Local storage backends for the two storage profiles:
- ```sh
- docker compose --profile postgres-vector up -d # Postgres + pgvector on 127.0.0.1:15433
- docker compose --profile postgres-qdrant up -d # adds Qdrant on 127.0.0.1:16333
- ```
- To run the Keep server itself with either profile, generated secrets and
- health-gated start-up, use `deploy/compose.yaml`; see
- [Run with Docker Compose](docs/server.md#run-with-docker-compose).
- ## Connecting coding agents
- `keep mcp` serves Memory and Knowledge to Claude Code, Codex, Copilot CLI and
- other MCP clients; `keep serve` exposes the same tools over HTTP on
- `127.0.0.1`. Agents that use the same project share one memory. The
- recommended setup is one `keep serve` daemon, with every agent connected
- through `keep mcp --server`:
- ```sh
- cargo install --path crates/keep-server --features local-embeddings
- keep --user me serve & # the daemon
- keep --project my-app mcp --server http://127.0.0.1:7357 # in each agent
- ```
- The daemon can hold Memory in memory or in Postgres, embeds Knowledge with
- a local ONNX model (no hosted model API unless you configure an
- OpenAI-compatible model for fact extraction and summaries), and supports
- scoped tokens, TLS, per-caller evidence budgets and per-source read
- allowlists. A container image is published at `ghcr.io/tulipfarm/keep`.
- [docs/server.md](docs/server.md) covers storage, scopes, the tool list,
- security and copy-paste client configuration.
- See [docs/benchmarks/methodology.md](docs/benchmarks/methodology.md) for how
- runs are designed, what the numbers mean and what they cannot tell you.
- ## Links
- - Website: <https://keep.tulipfarm.ai>
- - Source: <https://github.com/TulipFarm/keep>
- - Docs: [Server](docs/server.md), [Memory](docs/memory.md), [Knowledge](docs/knowledge.md), [Benchmark methodology](docs/benchmarks/methodology.md)
- - Built by [TulipFarm](https://tulipfarm.ai)
- ## Licence
- Apache-2.0. See [LICENSE](LICENSE). Benchmark datasets have their own
- licences (LoCoMo is CC BY-NC 4.0); the harness downloads them for local
- evaluation only.
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.