A research runtime for bounded decision models: it pages and refines a virtual option space so Jev can choose among far more tools, memories, and context blocks than a single request allows, with explicit approval for side effects.
Research design for a Jev-native agent system:enable more options than jev provided with virtulization and paging, tool integration, external helper logits Top-k proposals with Jev-controlled fallback ,decision-aware hierarchical memory, and dependency-aware replanning.
Pages a virtual option space so a bounded decision surface can exceed Jev's 255-option limit per request.
Transitions are PAGE, REFINE, CLARIFY, STOP, and context refresh, preserving stable IDs and revisions.
Core profile runs with no GPU, weights, or API key; :plan proposes and only :approve permits a side effect.
ExecutionReceipt marks a side effect applied, rejected, or unknown; unknown routes to needs_reconciliation, never an implicit retry.
The README reports 159 passing tests on the latest remote run — runtime and protocol checks, not an agent-quality claim.
Watch out
No license is shown in the digest, so terms are unclear. It is a research runtime: the README reports 159 protocol tests but warns that some live workloads are scripted, the core CLI is explicit rather than autonomous, and latency includes a noisy remote service.
Questions about this entry
Short answers drawn from the linked project or page.
Bounded decision models answer a limited number of options per request; the runtime pages and refines a larger logical option space over tools, memories, and context blocks while keeping the resident decision surface small.
At most 255, as the README notes; the runtime keeps the resident surface below that and reserves room for its own control choices like PAGE, REFINE, STOP, and CLARIFY.
Yes. The core profile runs locally without a GPU, model weights, or an API key and is a deterministic protocol and workspace smoke test, not a language-model quality benchmark.
:plan prepares a candidate action and :approve is the only command that permits a side effect, so workspace confinement, tool arguments, and structured traces stay inspectable before anything changes.
An ExecutionReceipt is applied, rejected, or unknown; an unknown receipt puts the task in needs_reconciliation until an external observer confirms the result, and it is never retried implicitly.
Only the optional 0.8B helper that proposes natural-language candidates or tool arguments does; install it with pip install -e '.[models]' and keep weights in a cache outside the repository, with CPU sufficient for functional checks.
It is research; the README reports 159 passing runtime and protocol tests but says they are not a claim of production agent quality, and live Jev experiments carry scripted workloads and noisy remote latency.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
Hacker News: jevmem keeps Claude Code project memory by asking Jev which turns are worth saving.
JevSpawn, a compositional policy, connects natural-language task specifications to finite probabilistic exploration: parallel action spawning pairs with feedback-driven branch selection, representation revision, and recovery from retained alternatives, with no training. It improves performance on eight benchmarks.
A free Claude Code skill that puts Jev in front of the expensive model: thirteen commands sort, check, score, gate and verify with Jev, so Claude reads only the unsure slice and writes the output. Needs python3 and an OpenRouter key.
An agent harness where Pi runs the loop, Gemini does the work, and Jev makes the small decisions: which model tier to use, which runbook sections to load, whether a tool call may run on a real Postgres database, and whether the report holds up.
A Python toolkit that hands browser sub-tasks to Jev instead of the coding agent: one fast_run call lets Jev drive clicks and typing step by step, and the agent gets back a typed outcome or a typed reason to stop.
A headless Python voice-assistant daemon for Arch/Hyprland desktops and macOS: middle-click summon, DeepSeek-Flash reasoning, xAI Grok voice, about two dozen desktop tools, and optional Jev triage over memory.
