LONG-TERM MEMORY FOR YOUR AI
Past conversations.Future understanding.
Give your AI memory across conversations, observations and actions. Recall the context. Trace it to the source.
“I take my tea without sugar.”
The cup slipped.
- Conversational memory
- Visual memory
- Action memory
- Cognitive memory
Your tea, without sugar. The blue cup was on the counter, but the last grasp slipped. Confirm its position before trying again.
- Interact
- Remember
- Recall
- Answer
BUILT FOR YOUR APPLICATION
One memory loop. Different applications.
INSIDE EMEMOX
One memory. Every connection.
Explore a record, compare its revisions, and test how it returns in another session.
DEVELOPER INTEGRATION
Write. Check. Recall.
Choose OpenAPI over HTTP or native ROS2 services. Both connect to the C++ memory engine: save experience, check processing, and recall it in another session.
Connect applications with a project API key over HTTP. Start with the OpenAPI contract.
Save dialogue, wait for extraction, then recall it in another conversation.
Keep the returned tracking_event_id. Reuse the same request_id and body on retry; use a new request_id for a new conversation input.
- YOUR SYSTEMPerception & interaction
Images, dialogue, state, action events
- EMEMOXMemory & cognition
Evidence, formation, recall, revision
- YOUR SYSTEMPlanning & execution
Use context, verify, act, report outcomes
THE PRODUCT
From experience to usable memory.
Build the memory loop through customer APIs, then inspect and manage it in the console. Explore the capabilities your application can use.
Experiences & knowledge
Customer API
- Conversations, observations and action attempts, linked to their evidence.
- Versioned materials and guidance that change with your application.
Recall & context
Customer API
- Bring relevant memories from earlier sessions into a new request.
- Combine words, meaning, entities, relationships and time.
Memory governance
Customer API & console
- Inspect the source and revision behind a selected memory.
- Correct, withdraw or expire content; export snapshots and read changes.
Developer platform
Signed-in customer console
- Connect through the API and test the memory loop in the Playground.
- Manage projects, scoped keys, team access, requests and usage.
How memory works
Follow a preference, a robot attempt or a cognitive snapshot from experience to later use.
THE MEMORY LIFECYCLE
Remember. Recall. Keep it up to date.
Carry an experience into a new conversation, revise it when things change, and inspect the evidence behind it.
Save the experience, with its source.
Keep the original message and its time. Once memory processing finishes, the preference can be recalled in later sessions.
An accepted event is not yet a completed memory. Check processing before using it.
EXAMPLE · A PREFERENCE OVER TIME
Alex · Drink preferenceOriginal message
“I like my tea without sugar.”
Conversation · Session 01 · 09:10
Alex prefers unsweetened tea.
- Source message
- “I like my tea without sugar.”
- Source session & time
- Session 01 · 09:10
Illustrative example · Local sample data, no live customer records.
Evaluation results
Memory retrieval quality, latency and context size, with supporting answer evaluation.
LoCoMo
1,986 evaluation questions
Memory retrieval
95.3%
Evidence Hit@10
- Turn recall
- 85.5%
- MRR
- 0.7615
- Mean retrieval latency
- 547.1 ms
- Mean retrieved context
- 26,098.9 bytes
Hit@10 counts questions with at least one labeled evidence match; turn recall measures evidence completeness.
Retrieval denominator: 1,982 / 1,986 questions; 4 without evidence labels.
Retrieved context is measured in bytes, not tokens.
Evidence hit rate by question type · Hit@10
Memory-assisted answering
Supporting evaluation of retrieved evidence with an answer model.
1,540 selected answer questions · 446 questions excluded from answer evaluation
68.1%
Answer correctness · LLM judged
- Mean generation latency · saved-evidence replay
- 2,839.5 ms
- Answer input tokens / selected question
- 17,723.7
Answer model: deepseek/deepseek-flash · Judge model: deepseek/deepseek-flash
Token values require reported usage for every selected question. Context bytes are not tokens.
Answer input coverage: 1540 / 1540; output coverage: 1540 / 1540. Output tokens / selected question: 234.8.
Answer correctness by question type
1,049 / 1,540 selected answers correct. 1,540 judged; generation failures: 0; judge failures: 0. Failures remain in the selected-question denominator.
LLM-judged hallucination flag rate: 31.4%. Denominator: 1,540 judged questions; lower is better.
Complete run · Quality checks failed
- Evidence retrieval stage
- Measured in this run · → UTC
- Answer generation and judging stage
- Measured in this run · → UTC
Completed ·
Run ID: 20261022-answer-recovery-v27-full
Source revision 30825f8a · Working tree has changes
Results apply to the recorded code, models, data and execution conditions. These dialogue-memory benchmarks do not measure robot perception or control.
TRY IT. BUILD WITH IT.
Try cross-session memory, then connect your app.
Try conversations, eight simulated scenes and cup actions in the demo, or start an API project for your own application.
Plan your integration.
The Free API plan has separate request allowances. Paid capacity is quoted and activated by the team.
View plans and allowancesDiscuss plans and integration