The memory runtime for agents that live longer than a context window.
Kept keeps everything an agent experiences, then returns only the part of its life worth loading now, with the original evidence attached.
Recall
Ask the demo agent about its past. Every bid you see is the one that made the choice.Inspect
The hierarchy is an index; every line opens down to the raw events.Forks
Same life, different memory. Four clones of one agent, fed the same future; only the memory policy differs.Developers
Send Kept everything your agent experiences. Ask it what matters now.from kept import Kept
memory = Kept(agent="my-agent", api_key=KEY)
memory.remember({"type": "tool_call",
"content": "Deployment failed with authentication error."})
context = memory.recall("What happened the last time deployment failed?",
budget=3000)
answer = agent.run(task=current_task, memory=context.text)
memory.credit(context, success=True, used=context.event_ids[:1])
SDKs, one file each with no dependencies: Python · JavaScript · MCP server (Claude and any MCP client) · tool definitions for OpenAI and Anthropic function calling. The demo agent answers recall and credit without a key; your own agents get a key each, and a key reaches only its own agent.
Keys are invite-only while Kept is in preview. Leave an email and one line on what your agent does, and a key will come to you.
Research
Generated lifetimes; the numbers are what the run printed.The benchmark is running.
event log · spine hierarchy · recall API · SDK · top-p routing · receipts · Memory Market · credit · learned routing · inspector · contradictions as chronology · workers · time questions · selective forgetting (hot, warm, cold, archived; pins) · dreaming (files untagged experience, new concepts, recurring failures, contradictions) · forks and clones (memory policy per agent; same future to every fork) · beliefs with versions and memory debt · ablation, forced recall and counterfactual credit · replay against an old self · poisoning and the immune system · surprise, obsession and false familiarity monitors · imports (jsonl, chat, git, terminal, slack) · Python, JavaScript and MCP integrations · encryption at rest · health watchdog
the 10,000-question study (spec 42) and the two-router divergence study (spec 46) haven't been run · learned routing hasn't shown a gain yet · the demo life is generated