Without memory
The model answers using only your question and its existing knowledge.
PERSISTENT MEMORY, ACROSS SESSIONS.
Persistent AI memory
from your own data.
Save a document, a policy, or a few facts as memory. Select it for later questions, or replace it when your knowledge changes. Compare the same prompt with and without memory.
Knowledge stays. Weights stay the same. Context is prepared once and saved as model attention state. Each memory is specific to its model.
Choose a model
One server. One inference task at a time.Connecting to the demo…
01
Build a memory
Keep the context outside your question.
Delete this saved memory? It will no longer be available for questions.
Your text, memory, and outputs are processed and stored on this server. Don’t submit secrets or personal information. Use the delete button to remove a saved memory. Usage counts and timings are recorded.
02
Ask. Compare. Explore.
One question, two runs of the same model.
Your request is queued
The difference, side by side
Actual model responsesWith your memory
The same model can also attend to the context you saved.
Both runs are processed in sequence. Response time depends on model size, context length, and current server load. This demo does not guarantee correct answers.
01 / PREPARE
Your context is converted into a saved, model-specific attention state.
02 / RESTORE
That state is loaded when you ask a question. No fine-tuning is performed.
03 / COMPARE
See the actual output with memory alongside a run without it.
MEASURED BEFORE THIS DEPLOYMENT
Local research results
Earlier tests used original BF16 models on an Apple M4 Pro with 48 GiB unified memory. These are synthetic, serial research workloads. The quantized CPU models above have their own behavior and performance.
Tacit-4B BF16 · open AnyJev-family decisions
48 synthetic policy questions42/48
correct with native memory
12/48 without the policy
2.83×
faster than fresh full-context input
219 ms native · 621 ms fresh full
Fresh full-policy input also scored 42/48. Timing used 144 requests per method. This compares saved native state with freshly processing the full policy, not an already-resident prefix cache.
Qwen3-14B BF16 · document answers
24 answerable facts22/24
facts recovered with native memory
0/24 without the document
4.47×
faster than fresh full-context input
0.96 s native · 4.28 s fresh full
Including unknown-answer questions: 30/32 overall, tied with full-document input and BM25 retrieval. Short-document timings used 24 requests per method.
Model loading and one-time memory preparation are excluded from these request timings. Storage was warm; the studies establish neither cold-storage performance nor general superiority in RAM use or over an already-resident prefix cache.
A result we keep visible
The accuracy improvement is uneven: ordinary queries improved 27/60 → 55/60, but counterfactual questions regressed 13/36 → 7/36. The optimized pipeline therefore remains optional. This is a separate harder typed-model evaluation, not the 48-question result above.
Longer documents, memory costs, and measurement limits
At roughly 24K document tokens, Qwen3-14B native memory averaged 6.67 s, versus 169.10 s for fresh full-document input. BM25 was faster at 3.06 s. Native memory, fresh full input, and BM25 each scored 9/12 on this correlated expansion test.
Those size-specific timings use only three requests per method. Mean peak MLX allocation was 35.061 GiB for native memory, 31.622 GiB for fresh full input, and 27.827 GiB for BM25. MLX allocation is not total process or whole-machine RAM. The evidence supports persistence and reuse; it does not establish a general speed/RAM improvement or replacement for all RAG and fine-tuning.