AtMem.ai
Your agents remember
only what you allow.
Control what enters agent memory and context. Review what happened in each run from your local dashboard.
Install locally
python -m pip install --upgrade atmem==2.3.10Current stableAtMem 2.3.10
DolphinBench · 600 tasks
Memory measured in actions,
not answers.
Remembering a fact is the start. Using it to send the right message, update a record or schedule a meeting is the test.
AtMem · task success
AtMem completed 390 of 600 tool-using tasks across three personas in its recorded DolphinBench run.
Each square represents 10 tasks. 390 passed; 210 did not.
- Mem070.7%424/600
- Hindsight69.5%417/600
- Honcho68.5%411/600
- Built-in memory65.7%394/600
- AtMemOur run65.0%390/600
- Supermemory59.2%355/600
Published references alongside our separate self-tested run—not a controlled head-to-head or an official AtMem ranking.
View the pinned reference data ↗How to read this benchmark
DolphinBench tests whether an agent applies remembered information in tool-using tasks. A task passes only when every required check passes. Each persona contributes 200 tasks.
AtMem was tested as 2.3.10b1, then promoted to stable 2.3.10 with version-only product changes. Reference rows use the same named Hermes/Luna reader, but are separate published runs; transports, ingestion and execution conditions are not established as identical. No matched comparative advantage or statistical significance is claimed.
Our submission is self-tested and awaiting maintainer review. Public-task subsets informed development; this is not an untouched holdout. The evidence card records configuration, ingestion provenance and the Morgan/129 transcript-recovery exception. Reference snapshot checked 2026-10-23.
Read the methodology and download the evidence ↗ · Follow the submission review ↗
Memory control
See what enters context and what stays out.
AtMem checks scope and policy before a memory reaches your agent.
Illustrative example · invented records
Admitted into context
Private spaceagent:research · workspace:acme
Customer prefers async updates over calls.
Renewal date moved to March.
EvidenceView the decision
- decision
- authorized · policy generation 41
- scope
- agent:research · workspace:acme · private
Private workspace scope
Kept out of context
WithheldSalary discussion from another workspace
Outside the requested workspace scope.
DecisionView the decision
- reason code
- scope_boundary_denied
- effect
- not delivered to the model
WithheldAn earlier note the user deleted
Deleted records do not enter candidate content.
Optional AtMem Code
One project. Shared memory across coding agents.
Find project knowledge, revisit earlier experience, and inspect what your agents remember.
Carry knowledge between agents.
Authorized coding agents can share project memory through the same project and shared profile. Private profiles remain separate.
Navigate project structure.
Explore source-backed structure and fetch selected code evidence on demand. Rich multi-language indexing uses an explicitly approved local worker.
Keep memory inspectable.
Search previous coding experience, inspect recorded outcomes, correct or delete memory, and move encrypted project bundles between clones.
Agent continuity
Preserve work. Resume with evidence.
A connected workflow can reuse a saved result after a restart, check a supported destination after a lost receipt, or stop an uncertain action for confirmation.
Keep completed work.
Installed document-workflow tests left one published document after crashes both before and after receipt storage.
Make uncertainty visible.
In the recorded public retail fault test, AtMem-connected runs stopped for confirmation instead of repeating the exchange request.
Inspect attempts in AtFlows.
Follow instrumented retries and reported costs in AtFlows. Group Timeline events by tool, service, model, type or hour, then click through to the records.
Continuity requires explicit integration with registered tools; it is not automatic for every agent. These are specific acceptance and development results, not a general task-success advantage. Repeated held-out evaluation remains open.
Control
Set memory boundaries.
Choose what an agent can store and retrieve within its scope.
Evidence
Review each run.
Inspect the context, model exchange, tool calls and recorded results.
Local
Run it on your machine.
Install AtMem, open the dashboard and connect a supported agent.
Open-source portfolio
Memory, intelligence, observability and supervision.
Choose the tools your workflow needs. AtMem governs memory; AtSys independently supervises coding tasks and recovery.
AtFlows
Local traces and cost visibility with optional AtMem-owned dashboard login.
AtFlows repository ↗AtSys
Progress monitoring, read-only diagnosis, permitted repair and independent task validation.
Watch AtSys and get started →Open source
Start with AtMem.
Explore the code, documentation and current release.