Queryable historical state
Separate session dates from event dates and preserve validity intervals, so past fact states remain queryable after updates.
Live Memory Playground: watch long-term memory form and change.
Natural language → two-stage LLM extraction → temporal memory → t4 answers and source evidence. Unavailable models produce an error, never a rule-based substitute.
While the backend wakes: watch a recorded real run
A recorded run of the Chinese fact-update preset, using real models on a local backend and fictional dialogue. Waiting is sped up; no audio. This is not current service status and uses no visitor model quota.
Recorded outcome: TensorFlow changes to PyTorch. Historical queries return TensorFlow before the update and PyTorch after it.
Transcript and recording details
Recording provenance and usageInteract with long-term memory
Connecting…
Presets start a fresh demo namespace and delete the current one. You can also enter your own natural language.
Visitors share model quota. Resume an interrupted preset from its unfinished step; wait for the countdown after a rate limit, or watch the recording above.
Conversation 1
Empty means today. This is when the conversation happened, not when a fact did.
Writes and questions send relevant text to the backend model. API keys stay on the server. A new conversation retains long-term memory without sending the old on-screen transcript.
Memory Inspector · facts and timeline
- Save a message to inspect extracted facts here.
Evidence behind this answer
Ask a question to inspect the sources of selected memories.
Historical state by date · as_of
Facts in force on a date · as_of
- Pick a date to see the facts in force at the end of that day. Reads stored intervals only; no model call.
Memories this answer did not use
Memories this answer did not use
- Shown after a question.
Answer request prompts, including raw evidence and fallback
HOW IT WORKS
Memory tracks change. Source text preserves detail.
Write long-term memory from past conversations, then retrieve relevant facts and source text for later questions.
WRITE
- Past conversationsSource text + session date
- Two-stage extractionExtract facts, then structure them
- Deduplication + timelineAdd, replace, terminate, coexist
- User-scoped storageSQLite / FTS5 + NumPy; memory + source text
READ
- A later questionCaller owns the current conversation
- Retrieve this user’s evidenceRelevant memory + selective source turns
- t4 answer contextFact state, source hydration + time notes
- Answer + evidence traceInspect the memory and source text used
Conditional path: the answerer requests source text or finds no evidence → search this user’s archived turns → answer again only if text is found; otherwise retain the verdict or return unknown.
BENCHMARK EVIDENCE
Concurrent regression and historical benchmarks, separately.
—
On narrow screens, scroll the table horizontally for the full comparison.
| Tested configuration | Correct per repeat | Mean accuracy | Recorded answer + grading tokens |
|---|
—
Naive RAG and Full Context were not rerun in this regression; it cannot establish parity with full history. No post-hoc significance claim is made. Token usage is neither a monetary bill nor total cost.
Loading results… data full report
BUILT FOR LONG-TERM MEMORY
Keep facts, time and provenance together.
Inspectable answer evidence
Inspect selected memories, source turns in the first prompt and conditional recovery. Provenance helps diagnose errors; it does not guarantee a correct answer.
User isolation + tool access
Memory and source reads stay within the user namespace. Integrate through REST, MCP, Python or CLI; the caller manages current conversation context.
- Python engineering tests
- —
- Line coverage
- —
- CI platforms
- —
—
Project report, known failures and negative results →