Agents run out of context. Long sessions fill with tool output, old messages and notes, and something has to decide what stays. A memory or context decision is a small yes or no, or a relevance score, asked many times. That suits a decision model, which returns a typed answer and a confidence value instead of a summary written in prose. The examples in this hub run on Jev and other decision models. Most are plugins for coding agents. compozy/yoshi prunes conversation history for Claude Code and Codex, kunchenguid/compact-adviser judges when it is safe to run /compact, and arslanr-com/perfectrecall decides what is worth recalling. Before you copy one, check what it removes and whether the removed text can be brought back. Check how it treats a case the model is unsure about, which model still sees the full context, and whether the author compared sessions with and without it. Entries differ a lot in how aggressive they are, from hints to fully automatic pruning.
15 entries
Showing 15 of 15
Claude Code plugin that replaces the compaction summary with Jev decisions. Every tool call and result is scored in one request against the whole conversation, stale ones are dropped or truncated, and everything kept stays verbatim. Also usable as an npm library.
Agent skills for Hermes, Claude Code and Codex that hand Jev the routing decisions: which model tier to use, which search results to open, which memory passages to read and which hide injected instructions. Ships a dashboard with an on, shadow and off switch and publishes measured latencies per skill.
A Claude Code and pi plugin that uses Jev to judge whether a session has reached a safe point to run /compact, in a hint mode or a fully automatic mode.
Kun Chen (@kunchenguid)Openkunchenguid/compact-adviser on github.com183 starsEval set 40 sessionsHint floor (empty) 0.90
A Claude Code plugin that uses Jev to trim noisy Bash output after a command runs but before the result reaches the main model, once the output passes an estimated token threshold.
A Claude Code hook that splits large tool results into blocks and asks Jev a yes/no relevance question per block, replacing low-relevance blocks with a short recoverable stub before they enter context.
Local context-pruning proxy for Claude Code and Codex. Jev judges which conversation history is still needed above a configurable size gate, and Yoshi applies the validated omissions before forwarding to Anthropic or OpenAI; the author calls it a proof of concept headed into CompozyOS.
An asynchronous Jev integration for Hermes Agent that supervises turn admission, tool-call gating and context curation in the background, using up to sixteen typed choice, score and noul questions per call.
Pi coding-agent extension that splits large tool results into blocks, has Jev score whether the agent will need each one, and hides low-scoring blocks behind a stub recallable with pi_jev_recall.
Eugene TrapeznikovOpenpi-jev-triage on github.com14 starsContext freed 9.7%Judge spend $0.65
An omp plugin that scores every old tool output with Jev, keeps what is still needed, and parks the rest on disk, cutting context by 29 to 55 percent across the author's own sessions at about $0.0005 per scoring pass.
Filters every tool result through Jev before the coding model reads it, with an A/B harness that measures pass@1 and cost against an unfiltered control.
A pi agent extension that curates the context sent to the LLM before each call, using a one-question-per-unit Jev judgment to decide which prior messages are still needed.
Agent runtime in Rust where the context, tools, permissions and requests are explicit state you can read, change and put back, plus an MCP bridge and the kamchatka terminal agent. The linked example uses Jev for content-aware compaction in that agent.
A Codex plugin that asks Jev which bulky tool results a session still needs, then replaces the rest with a bounded head and a one-line note, so long sessions don't resend stale output on every turn.
An automatic extension for the Pi coding agent that has Jev score older tool outputs and stops resending the ones no longer needed, while keeping the original session messages intact.
Agent memory where Jev decides what is worth recalling, with Mnemosyne-compatible storage and a Hermes integration.
Nothing matches those filters
Clear a filter to widen the search, or send us the thing you were looking for.
The common loop is score and prune. A large tool result is split into blocks. The decision model gets one relevance question per block, such as whether the agent will still need it. Low scorers are replaced with a short stub. GhalebDweikat/winnow does this as a Claude Code hook and keeps the stub recoverable. Because the question is the same for every block, one batched call can cover a whole result.
Noul and score cover most of it. Choice appears when the question has named outcomes, such as keep or drop. tamaratran/fast-jev-compaction goes further than pruning: it replaces the compaction summary with decisions, scoring every tool call and result in one request against the whole conversation. Most of these entries run inside a coding agent, as a hook or plugin, so the pruning happens without the main model being asked to do it.
The risk is dropping something the agent needs later. Entries handle it by keeping the originals, as omp-jev-compaction does by parking old outputs on disk, or by acting only above a size threshold. Whether the saved context outweighs the occasional miss depends on your sessions, so test on your own. Start in a hint mode if the entry offers one, and read what it would have removed before you let it act.
Yes. Entries such as arslanr-com/perfectrecall and GhalebDweikat/winnow ask Jev whether a memory or tool result is worth keeping and drop or stub the rest. The decision is a typed score or yes or no that code acts on.
No. RAG retrieves documents to answer a question. Agent memory entries decide what an agent keeps in its own context or recalls from past sessions. The two overlap when an agent searches its own history.
It can. The better entries keep the original text on disk or leave a stub that can be expanded, and some only act once the context passes a size limit.
