fraise

Fraise

Docs · Discord · Query language · Issues

Fraise is a memory database for AI agents. One they query directly, in a language built for tokens, not humans.

remember 'acme moved to annual billing' topic:billing entity:acme

recall billing entity:acme since:30d top:5

Two verbs. One binary. No infrastructure to run.

Install

brew install fraisehq/tap/fraise
brew services start fraise
curl -X POST localhost:9876/api/v1/q -H 'content-type: application/json' \
  -d "{\"query\": \"remember 'the parrot is turquoise' topic:color\"}"

Linux packages, Docker, go install and signed release binaries are all in Get Started below.

Read Durability before you deploy. Persistence is issue #171 and the next major piece of work.

How it compares

Measured on LoCoMo — 10 multi-session conversations, 1,986 questions — with every system ingesting the same conversations, using the same extraction model and the same embedding model, and answering the same questions. k=10, full data, latest run per system as of 2026-10-11.

system	retrieval recall	p50 search	memory tokens
fraise 0.3.3	0.910	0.185 s	7,453,109
hindsight 0.9.2	0.913	5.836 s	1,688,873
everos 1.2.3	0.897	0.376 s	10,030,966
letta 0.16.8	0.897	0.349 s	226,854
mem0 2.0.18	0.886	0.576 s	4,120,078
memos 2.0.33	0.865	0.307 s	4,000,552
graphiti 0.30.2	0.828	0.290 s	9,861,973
cognee 1.5.3	0.790	6.273 s	8,526,019
langmem 0.0.30	0.783	0.227 s	1,078,110

Fraise comes within 0.003 of the best recall, 31× faster than the one system that reaches it. Near enough the same evidence, in a fraction of the time. The token column is the price of that recall: the extractor (fraise-sdk 0.2.0) reads each message beside what the graph already holds, and that context is where the recall came from.

These come from agent-memory-benchmarks, an open harness — precision, recall and F1 across k ∈ {1, 3, 5, 10}, per category, every run tagged and reproducible from the tag. The full tables are in its RESULTS.md.

Why Fraise

• A query language agents can actually write. FQL has two verbs — remember and recall — and one way to say each thing. Fewer degrees of freedom means fewer ways for a model to get it wrong, and fewer tokens spent saying it.
• Hybrid retrieval. Facts are indexed for full-text, graph, and (optionally) vector search. One query, ranked across all three.
• Temporal by default. Recent memories outrank older ones, so recall is recency-aware without asking for it.
• The fastest system measured. 0.185 s p50 on LoCoMo, ahead of every other system and 31× faster than the one with matching recall. Remember and recall mid-step, while the user waits.
• No infrastructure. A single binary. No database to provision, no service to stand up beside it.
• Open source, MIT.

Status

v0.3.3. The core loop works end to end and the install paths are verified on clean machines. The benchmark row above was produced from this release. 0.4.0 (late October) changes the scorer and adds explain and describe; 0.5.0 (November) is persistence and forget.

Good for building agent memory today. Not yet for long term production use.

How it works

Fraise stores knowledge as a temporal memory graph built from three kinds of node:

• facts — the things you remember, one statement each
• entities — who or what a fact mentions
• topics — what a fact is about

Edges connect facts to the entities they mention and the topics they're about, so a query can start from either side. A recall finds seed facts by text and, optionally, by vector similarity; a recall that names only topics or entities starts from the facts filed under them. With depth:1 or depth:2, a named topic or entity whose facts matched better than its size predicts passes that evidence on to the rest of its facts. Results are ranked by relevance and recency, and the best top come back.

Ranking is not a black box: a fact's score is its own match strength plus what it receives through anchors carrying more mass than their size would predict. The background rate that "more" is measured against is estimated per query, from the part of the graph the query touched, so there is no relevance constant to tune. Hybrid retrieval has the full model.

A single Fraise instance holds several independent memory graphs (8 by default), addressed with @N — one per user, per session, per agent, however you like.

Get Started

Fraise is a single binary — no database to provision, nothing to configure. Every route below leaves you with a server listening on port 9876, at localhost:9876.

With Homebrew

brew install fraisehq/tap/fraise
brew services start fraise

The service survives crashes and restarts on login (keep_alive), logs to $(brew --prefix)/var/log/fraise.log, and reads its config from $(brew --prefix)/etc/fraise/fraise.config.toml — installed with every setting commented at its default, and never overwritten on upgrade.

With Docker

docker run -p 127.0.0.1:9876:9876 ghcr.io/fraisehq/fraise:latest

Published tags: one per release, latest for the newest stable, edge for the tip of main, and an immutable full-commit-SHA tag for every merge that touches the server.

Images are built with SLSA provenance, verifiable without pulling:

gh attestation verify oci://ghcr.io/fraisehq/fraise:latest --repo FraiseHQ/fraise

With Go

go install github.com/FraiseHQ/fraise/cmd/server@latest
"$(go env GOPATH)/bin/server"

The binary installs as server, after its package path — rename it to fraise if that reads better.

Nothing further is needed to trust this: the Go toolchain checks every module download against the public checksum transparency log, and your own machine compiles the result.

Linux packages

.deb and .rpm packages ship with every release, with a systemd user unit:

VERSION=0.3.0
ARCH=$(uname -m | sed 's/x86_64/amd64/;s/aarch64/arm64/')
curl -sSfLO "https://github.com/FraiseHQ/fraise/releases/download/v${VERSION}/fraise_${VERSION}_linux_${ARCH}.deb"
sudo dpkg -i "fraise_${VERSION}_linux_${ARCH}.deb"
systemctl --user enable --now fraise

Logs go to the journal (journalctl --user -u fraise -f), and the unit reads ~/.config/fraise/fraise.config.toml when present — a shipped default with every setting commented lives at /etc/fraise/fraise.config.toml to copy from. For agents that outlive your login session, let the user manager keep running: loginctl enable-linger $USER. A system-level (shared server) variant of the unit is described in Running as a managed service.

From a release binary

VERSION=0.3.0
OS=$(uname -s | tr '[:upper:]' '[:lower:]')                # linux | darwin
ARCH=$(uname -m | sed 's/x86_64/amd64/;s/aarch64/arm64/')  # amd64 | arm64
ASSET="fraise_${VERSION}_${OS}_${ARCH}.tar.gz"
BASE="https://github.com/FraiseHQ/fraise/releases/download/v${VERSION}"

curl -sSfLO "${BASE}/${ASSET}"
tar xzf "$ASSET"
./fraise

Windows builds ship as .zip under the same naming scheme.

Verify a release

Releases carry a cosign signature over checksums.txt, using the same VERSION and BASE as above:

curl -sSfLO "${BASE}/checksums.txt"
curl -sSfLO "${BASE}/checksums.txt.sigstore.json"

# 1. the bundle proves checksums.txt came from this repo's release workflow
cosign verify-blob \
  --bundle checksums.txt.sigstore.json \
  --certificate-identity-regexp 'https://github.com/FraiseHQ/fraise/.github/workflows/go.yaml@refs/tags/v.*' \
  --certificate-oidc-issuer https://token.actions.githubusercontent.com \
  checksums.txt

# 2. checksums.txt proves your archive is the one it covers
sha256sum --ignore-missing -c checksums.txt   # macOS: shasum -a 256 --ignore-missing -c

Run from source

git clone https://github.com/FraiseHQ/fraise
cd fraise
make dev

Your first memory

curl -X POST localhost:9876/api/v1/q \
  -H 'content-type: application/json' \
  -d "{\"query\": \"remember 'the parrot is turquoise' topic:color\"}"

curl -X POST localhost:9876/api/v1/q \
  -H 'content-type: application/json' \
  -d '{"query":"recall parrot"}'

{
  "results": {
    "count": 1,
    "hits": [
      { "value": "the parrot is turquoise", "timestamp": "...", "score": 0.2867 }
    ]
  }
}

Use it from an agent

MCP

fraise mcp is a stdio MCP server — a thin bridge to a running daemon, so any MCP client can remember and recall. It exposes two tools, recall and remember, and needs no flags when the daemon is on its default address.

Start the daemon first (brew services start fraise, or systemctl --user start fraise on Linux), then register the bridge with whichever coding agent you use.

Claude Code

claude mcp add fraise -- fraise mcp

That registers it for the current project; add --scope user to make it available in every project instead.

Codex

codex mcp add fraise -- fraise mcp

Or write it into ~/.codex/config.toml directly:

[mcp_servers.fraise]
command = "fraise"
args = ["mcp"]

OpenCode

Add it to opencode.json in your project root:

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "fraise": {
      "type": "local",
      "command": ["fraise", "mcp"]
    }
  }
}

Any other client

{
  "mcpServers": {
    "fraise": { "command": "fraise", "args": ["mcp"] }
  }
}

The bridge describes itself over MCP: each tool arrives with a description and a full JSON schema for its arguments and results — the FQL shapes, a worked example, what a score means — so a client knows what recall and remember do and how to call them without being told. What a tool description cannot carry is the policy: the habit of reaching for memory unprompted, and the judgement about what is worth keeping. That belongs in the file your agent already reads — CLAUDE.md for Claude Code, AGENTS.md for Codex and OpenCode — and two habits are enough: recall before answering anything that leans on earlier decisions or preferences, and remember only facts that will still matter in a later session, one self-contained fact per call with the topics and entities that will make it findable.

SDKs

Python (sdk/python) — the only SDK today:

pip install --pre fraise-sdk

from fraise_sdk import FraiseClient

with FraiseClient("http://localhost:9876") as fraise:
    fraise.remember("the parrot is turquoise", topics=["color"])
    for hit in fraise.recall("parrot", top=5):
        print(hit.value, hit.score)

The Python SDK is dependency-light and supports vector search when you supply an embedder. See its README for embeddings and the full API.

TypeScript — not available yet (#179). Until it lands, TypeScript callers use fraise mcp or talk to the HTTP endpoint directly; it is two verbs over one route, so a client is a short wrapper around fetch. See the HTTP API.

Claude Agent SDK

The Python SDK ships memory tools for the Claude Agent SDK, exposed as an in-process MCP server so the agent decides what to store and recall:

from claude_agent_sdk import ClaudeAgentOptions
from fraise_sdk import FraiseClient
from fraise_sdk.integrations.claude_agents import memory_server, allowed_tools

fraise = FraiseClient("http://localhost:9876")
options = ClaudeAgentOptions(
    system_prompt="Remember durable facts the user shares, and recall them when relevant.",
    mcp_servers={"fraise_memory": memory_server(fraise)},
    allowed_tools=allowed_tools(),
)

A complete, Docker-runnable agent lives in examples/claude-agent-sdk.

OpenAI Agents SDK

memory_tools(client) returns bound recall and remember tools:

from agents import Agent, Runner
from fraise_sdk import FraiseClient
from fraise_sdk.integrations.openai_agents import memory_tools

fraise = FraiseClient("http://localhost:9876")
agent = Agent(
    name="Assistant",
    instructions="Remember durable facts the user shares, and recall them when relevant.",
    tools=memory_tools(fraise),
)

result = Runner.run_sync(agent, "My favourite colour is orange.")
print(result.final_output)

A complete, Docker-runnable agent lives in examples/openai-agents.

References

• Roadmap
• Documentation
• Architecture
• Query language grammar
• Release process
• Issues

Contributing

Contributions are welcome — see CONTRIBUTING.md for how to build, test, and submit changes.

Code of Conduct

This project follows the Contributor Covenant.

Community

Questions, ideas, or building something with Fraise? Join the Discord. Bugs and feature requests belong in issues so they don't get lost.

License

MIT — see LICENSE.