Your agents don’t need better memories.
They need your team’s context.
Context management for AI agents — every decision, constraint and lesson, kept current and handed to whoever works next
Plate I — the site
one shared workplace · claude code · cursor · codex
surveyed 2026·07·28
claude mcp add --transport http klio https://mcp.klio.tech/mcp \
--header "X-Klio-Key: YOUR_KEY" \
--header "X-Klio-Agent: claude-code"Works with Claude Code, Cursor, Codex and any MCP client. Free for one person — unlimited memories, kept forever.
Plate II — the condition
Every agent keeps a private notebook.
Each vendor built memory for its own agent, inside its own product. Your Claude context stays in Claude. Your editor’s context stays in your editor. No vendor has any reason to hand your work to a competitor’s agent.
So you become the integration layer.
Observation i
You re-explain the same decisions, every session, to every tool.
Observation ii
They contradict each other. One agent quietly undoes what another just decided.
Observation iii
You pay twice — tokens and hours — re-examining work that was already done.
The re-explaining is the symptom.
Rework is the bill.
Plate III — the ledger · repo/api · shift 14:00–15:00
Work belongs to the job, not to the tool that happened to do it.
Nothing was re-explained. Three agents, three vendors, one job — each opened the same workplace and continued from the last entry.
Plate IV — the apparatus · the handover loop
Clock in
Any agent connects over MCP. Claude Code, Cursor, Codex, your own. No vendor owns the door.
Capture
Work is recorded as it happens. Secrets and personal data are stripped before anything is stored.
File to the job
Everything lands under a project. The job's knowledge stays with the job, not in one global pile.
Distil
Raw activity becomes the few durable facts worth keeping — the next agent inherits conclusions, not transcripts.
Retire
When a newer decision contradicts an older one, the old one is superseded. Nobody acts on stale truth.
Hand over
The next agent picks up current state and keeps going. One writes; the next to look already sees it.
Movements iv and v are the ones a plain store cannot perform. Anyone can keep text. Almost nobody retires it when it stops being true.
The next agent starts where the last one finished.
Plate V — the site at shift change
Plate V·B — the instruments · seven verbs
Seven MCP tools your agents already know how to call.
recall(query)Pull the relevant prior context before acting.remember(fact)Persist something worth keeping across sessions.observe(event)Log what happened — raw session activity.plan(intent)Record the plan so the next agent builds on it.decide(choice)Capture a decision so it is not re-litigated.note(text)Jot a durable note scoped to the project.space(name)Open or switch a scoped memory store.No SDK to wire, no framework lock-in. Any MCP client gets these verbs out of the box.
Plate VI — the specification · scope & refusals
Strong scope is the feature. Random global memory everywhere is the failure.
Kept
- Decisions — what was chosen, and why
- Repo rules — conventions the job must hold to
- Plans — the shape of work in flight
- Durable facts about the project
Never kept
- Secrets, keys, tokens — stripped before storage
- Personal data — redacted at the door
- Transcripts, tool calls, file reads
- Anything a newer decision has retired
Scoped to
- A project — the job's knowledge stays with the job
- A team — shared by the people on it, not the internet
- An agent identity — you can see who wrote what
Open to
- Any MCP client — Claude Code, Cursor, Codex, yours
- Any vendor — we are not one of them
- Your own hardware — the engine is open source
Redaction runs before anything is written, and fails closed. If it cannot be made safe, it is not stored.
Plate VI·B — the comparison · the shelf
Most memory tools recall for one agent. Klio is the shared workplace.
| Capability | Klio | mem0 | Zep | Supermemory |
|---|---|---|---|---|
| Cross-agent shared memory | ||||
| Local-first (runs on your machine) | ||||
| Encrypted under a user-owned key (self-hosted) | ||||
| MCP-native | ||||
| Open source |
Plate VII — the schedule · rates
Free for one pair of hands. Paid when the shift has a crew.
Free
$0
forever
- People sharing a brain — Just you
- Sync — Across your own devices
- Artifacts — 100 MB
- Memory retention — 30 days
Team
$20
per user / month · $200 yearly
- People sharing a brain — Everyone you invite
- Sync — Real-time, across the team
- Artifacts — 5 GB pool
- Memory retention — Forever
Platform · Enterprise
Custom
per end-user
- Embed Klio in your own product
- Isolated memory — per end-user
- SSO — audit trail — DPA
- Support — a person, not a queue
Connected agents and memories are unlimited on every tier — you pay for people, not volume. Retention is 30 days on Free and forever on Team. Self-hosting the open-source engine on your own hardware is free — the source is at github.com/klio-tech/klio. The full schedule is published at klio.tech/pricing.
Appendix — questions put to the surveyor
What is Klio?
Klio is a memory layer for AI agents. It captures what your agents learn, stores it, and serves it back through MCP so your agents remember across sessions and share context across tools — which is what keeps them reliable. Self-hosted, that memory is encrypted at rest under a key you own.
Does my data leave my machine?
No. Klio is local-first: self-hosted, memory is stored on your machine and encrypted at rest under a key you own. Nothing leaves unless you explicitly opt into Klio Cloud, where memory lives on our infrastructure under keys we manage.
How is Klio different from mem0, Zep, or observability tools?
Memory tools like mem0 and Zep recall for a single agent. Observability tools watch output after the agent acts. Klio gives multiple agents a shared, persistent memory — preventing drift at the source instead of recalling for just one agent or flagging problems after the fact.
How long does it take to set up?
One command: `npx @klio-tech/klio init`. It wires Klio into Claude Code, Cursor, or Codex and starts remembering from the next session.
Which agents does Klio work with?
Klio is MCP-native, so it works with Claude Code, Cursor, Codex, and any MCP client out of the box.
Can I change the embedding model?
Yes. Embeddings are pluggable per space — nomic-embed-text by default, with snowflake-arctic-embed2, OpenAI text-embedding-3-small, and others supported. Pick the model that fits each store.
Does Klio run on Windows?
The self-hosted engine runs anywhere Docker does, including Windows via WSL2. If you’d rather not run anything locally, Klio Cloud works from any OS with zero install.
Is it really encrypted?
It depends where you run it, so here is the honest split. Self-hosted: memory is encrypted at rest under a key you own, and every write is chained with SHA-256, so the history is tamper-evident and inspectable. Klio Cloud: memory is encrypted in transit over TLS and at rest at the infrastructure level, secrets and PII are redacted before storage, and every org is isolated — but the keys are ours, not yours, and Cloud writes are not hash-chained today.
Plate VIII — the workplace for AI agents
Your agents don’t need better memories. They need somewhere to work together.
three decisions, one rule
three decisions, one rule
no work paid for twice