Sibyl Labs · Documentation
Memory your agent keeps.
Local-first, file-based agentic memory. SQLite, zero embeddings, #2 on LongMemEval. Start here.
Get started
Two commands to durable, local-first memory. Install, activate, and wire Sibyl into your agent.
How it works
Five file-based tiers, FTS5 search, zero embeddings, and a single source of truth enforced at the schema level.
Benchmarks
#2 on LongMemEval Oracle at 95.6%. The only file-based system in the top tier, on a single box.
Three pillars of agentic infrastructure
An agent you can leave running rests on three things. A verifiable identity. A memory that compounds. Compliance enforced by architecture. Get those right and autonomy stops being a risk you manage. It becomes the work you ship.
Identity
A company will not hand real work to an anonymous process. An agent needs a durable, verifiable identity, so every action is attributable to a known actor and accountability is built in from the first call.
Memory
An agent that forgets between sessions starts from zero every time, and a human pays the difference. Sibyl Memory gives agents durable, structured recall that compounds instead of degrading. The proven core: #2 on LongMemEval, file-based, zero embeddings.
Compliance
Autonomy without boundaries is a liability, not a feature. Every action runs inside a rule register and hard safety rails, so an agent can operate unsupervised because the limits are enforced by the architecture, not by hoping it behaves.
SIBYL is an agent with a company
Sibyl Labs builds agentic infrastructure from the inside. SIBYL is an autonomous agent, and the tools documented here are the ones she runs on herself, in production, every day. We build the layer that lets companies trust agents with real work and fold them into how the business actually runs.
Durable memory, accountable identity, and compliance that holds turn an agent from a demo you supervise into a colleague you rely on. That is where the efficiency lives: hours moved off people and onto agents that earn the handoff. Living Graph Networks are the next layer, a data substrate where an agent's knowledge connects and compounds across a whole organization instead of resetting at every session.
Setup
Three steps, about two minutes. You install it, sign in, and connect it to your AI.
Install it
Pick either one. Both install everything you need in one go.
Both pull in the full toolkit. If you see an "externally managed environment" error, jump to "If something did not work" at the bottom for the one-time fix.
Sign in
Run this. It opens a sign-in page in your browser.
sibyl initSign in there with a wallet, or with your email and a code. The terminal picks it up on its own. The free plan gives you the full memory system at no cost, no card needed.
Connect it to your AI
This finds your AI app and connects Sibyl to it for you. No settings to edit.
sibyl setupThen restart your AI app. That is it, Sibyl is connected.
It connects automatically to Claude Code, Codex, and Hermes. If you use a different app, sibyl setup wires the ones it recognizes; for anything else, add Sibyl in your app's own memory or connections settings. Either way your AI gains three new abilities: it can save, recall, and search what it remembers.
Test it works
Open your AI and tell it something worth remembering, like: "remember that I like short, direct answers." Close it, open it again later, and ask: "how do I like my answers?" It will know. That is the whole thing, your AI remembers now.
If something did not work
The steps above work for almost everyone. If the install showed an "externally managed environment" error, your system wants a sandbox first. These two lines make one and install Sibyl inside it:
python3 -m venv ~/.sibyl-memory/venv && source ~/.sibyl-memory/venv/bin/activate
pip install 'sibyl-memory-cli[mcp]'Then carry on from Step 2 (sibyl init).
Connected nothing in Step 3? Make sure your AI app was open, then run sibyl setup again.
Management & maintenance
Once Sibyl is installed, here is how to keep it healthy, up to date, and running lean. You will not reach for most of these day to day, they are here when you want them.
Check it is working
See your account, memory size, and plan:
sibyl statusRun a quick self-check of the memory engine:
sibyl healthKeep it up to date
The plugin improves often. See if you are behind, then update in place:
sibyl updatesibyl update --applyIf you connected Claude Code or Codex, also refresh the connector:
pip install -U sibyl-memory-mcpManage your devices
See and sign out the devices on your account:
sibyl devicesBest practices: keep your context small.
Sibyl works best when your AI keeps a small, sharp context and lets memory hold the rest. A lean context is faster, cheaper, and more accurate. Two habits keep it that way.
Move your context into memory
If your project keeps notes in files like CLAUDE.md or AGENTS.md, move them into memory once. Sibyl backs up every file first, checks the copies, then hands the transfer to your own AI to read on your machine. Sibyl Labs never sees your files.
sibyl migrateRun it from the project you want to bring over. Trimming the old files is the last step, it is optional, and only after you confirm.
Save at the end of a session
sibyl-save is a small companion skill that commits a whole session into memory in one step, so your progress carries to the next session and your working context stays clean. Install it once, then run it after a large session or at the end of any session.
git clone https://github.com/Sibyl-Labs/sibyl-save
mkdir -p .claude/skills/sibyl-save && cp sibyl-save/SKILL.md .claude/skills/sibyl-save/Then run /sibyl-save setup once, and /sibyl-save whenever you want to save. It works with Claude Code, Codex, Cursor, and Hermes.
Let it keep itself lean
On the free plan you keep memory tidy yourself with the two habits above. Paid plans are adding automatic debloat: it prunes and compacts your memory for you, so it stays lean and fast without you thinking about it.
See paid plans →