Paragraph library
Sample memory · Orbit
OPEN SOURCE · POWERED BY JEV
The right context.
For this moment.
Your agent doesn’t need every paragraph.
Give it the ones that matter.
Codex · Claude Code · Hermes · Your agent · Codex · Claude Code · Hermes · Your agent ·
THE LIVE PLAYGROUND
Ask differently.
Watch context change.
A fictional team. Twelve paragraphs of memory.
Every relevance judgment comes from Jev.
Choose a request to see real model selection.
How long did retrieval take?
12
Dynamic context
Only what the request needs
[ ]
A little less noise.
A lot more focus.
Run a request. Selected paragraphs appear here, with their sources intact.
Inspect the agent payload
No selection yet.
Relevance is Jev’s estimated probability that a paragraph helps with this request. The cutoff is adjustable, not a calibrated guarantee. Budgeting preserves whole paragraphs and counts the complete payload using o200k_base; other models may tokenize it differently.
MEASURED ON EIGHT EXISTING SITES
Less context.
Measured, not assumed.
Exact archived website requests, replayed through Jev.
Default relevance cutoff: 0.50 · September 20, 2026.
14.8%fewer generation-input tokens
230,057 → 195,953
34,104 tokens removed across eight requests.
All 80 required-paragraph checks passed.
Fewer generation tokens ≠ the same reduction in cost.
After Jev’s selection cost, estimated uncached input cost fell from $0.17254 to $0.16785: 2.7% savings. With cached generation prompts, the dynamic path cost more. This replay did not regenerate sites or establish equivalent output quality.
View exact token counts and test sites
| Test site | Full input | Dynamic input | Reduction |
|---|---|---|---|
| Trade negotiation matrix | 28,716 | 24,795 | 13.7% |
| Forecast ensemble lab | 28,777 | 20,948 | 27.2% |
| Skincare routine auditor | 28,994 | 23,294 | 19.7% |
| Command deck melder | 28,963 | 25,843 | 10.8% |
| Documentary release risk modeler | 28,716 | 24,814 | 13.6% |
| Tribute concert setlist planner | 28,484 | 25,825 | 9.3% |
| Prosthetic chair optimizer | 28,799 | 25,232 | 12.4% |
| Vehicle packaging studio | 28,608 | 25,202 | 11.9% |
| All eight sites | 230,057 | 195,953 | 14.8% |
QUESTIONS, ANSWERS, AND EVIDENCE
Does less context
keep the answer right?
29 questions in a revised subset. One answer model. Three independent arms.
Public text from the eight sites plus the demo’s sample memory.
29/29dynamic-context answers correct
96.1% less answer-model input
Full context: 29/29 (100.0%) · No context: 9/29 (31.0%)
Accuracy change: +0.0 percentage points.
Answerable questions: 20/20 · Missing-information checks: 9/9.
Fresh rerun of the revised subset.
These questions use 71 paragraphs of published site text and sample memory, not the full website-generation prompt above. The compression rates are not interchangeable. Three questions were excluded after failing the original run; this is not evidence of improved accuracy on the original workload. No dynamic answers failed this run; a small sample does not prove equivalence. One model, one fresh run, 29 hand-authored questions; this does not establish regenerated-site quality.
Compare accuracy and input saved by source
| Source | Full accuracy | Dynamic accuracy | Input removed |
|---|---|---|---|
| Command deck | 3/3 (100.0%) | 3/3 (100.0%) | 96.4% |
| Documentary | 3/3 (100.0%) | 3/3 (100.0%) | 96.2% |
| Forecast | 3/3 (100.0%) | 3/3 (100.0%) | 95.7% |
| Orbit | 8/8 (100.0%) | 8/8 (100.0%) | 96.4% |
| Prosthetic | 2/2 (100.0%) | 2/2 (100.0%) | 95.1% |
| Skincare | 3/3 (100.0%) | 3/3 (100.0%) | 94.9% |
| Trade | 3/3 (100.0%) | 3/3 (100.0%) | 96.1% |
| Tribute | 2/2 (100.0%) | 2/2 (100.0%) | 96.9% |
| Vehicle | 2/2 (100.0%) | 2/2 (100.0%) | 97.1% |
MEASURED WITH LIVE JEV CALLS
Parallel selection.
Measured overhead.
81 cold queries. Three corpora. One, three, and eight workers.
Same paragraph coverage, request, batch size, and relevance policy.
0.76smedian website-prompt retrieval
5.0× faster than sequential
3.82s with one worker → 0.76s with eight.
2.0× faster than the previous three-worker default.
Parallel does not mean zero overhead.
Across 12 paired questions, median complete-answer time was 3.03s with full context and 2.90s with dynamic context, including Jev. P95 was 3.56s versus 11.26s. The median paired difference was +3 ms; 6/12 dynamic requests took longer. Exact accuracy was 12/12 versus 11/12. This is a small measured sample, not a latency guarantee.
Immediate cache repeats made no new Jev calls: median 1.1 ms on the question corpus and 52.4 ms on the website prompt. Local context assembly still takes time.
Inspect median and tail retrieval times
| Corpus | Workers | Median | P95 |
|---|---|---|---|
| Demo memory · 12 paragraphs | 1 | 476 ms | 722 ms |
| Demo memory · 12 paragraphs | 3 | 452 ms | 775 ms |
| Demo memory · 12 paragraphs | 8 | 489 ms | 506 ms |
| Question corpus · 71 paragraphs | 1 | 2994 ms | 3740 ms |
| Question corpus · 71 paragraphs | 3 | 1022 ms | 1166 ms |
| Question corpus · 71 paragraphs | 8 | 566 ms | 784 ms |
| Website prompt · 78 paragraphs | 1 | 3819 ms | 5448 ms |
| Website prompt · 78 paragraphs | 3 | 1514 ms | 1872 ms |
| Website prompt · 78 paragraphs | 8 | 763 ms | 1280 ms |
A memory that adapts.
Without rewriting
what you know.
Your paragraphs.
Your database.
Import Markdown or text into a local SQLite collection. Keep exact words and source references. Replace a document to update its memory.
Meaning over
matching words.
Jev judges each paragraph against the full request. A shared keyword isn’t enough. Useful constraints and dependencies belong in context, too.
A focused context.
A clear provenance.
Relevant paragraphs become a source-tagged payload within your chosen budget. Your agent uses it as reference material for the task.
ONE ENGINE. YOUR AGENT.
Make memory
part of the workflow.
Download the self-contained skill or install the Python CLI.
Use your TypeSafe key. Keep the database on your machine.
The skill runs retrieval when invoked. It does not silently intercept every agent request. Add the supplied instruction to your agent’s project guidance for consistent use.
Read the setup guide ↗ · Download skill ZIP ↓