Sivtr Memory
SkillDocs & knowledgeRetrieve shared local work memory: terminal activity, AI conversation history, prior decisions, validation evidence, debugging trails, recaps, and handoff context. Use before asking the user to repeat local work context.
Available today. Use it from your connected AI after setup.
No other account needed.
Connect ahel once, and every AI you use reads what you have installed.
Then ask your AI: use the Sivtr Memory skill
What this skill tells your AI
The instructions your AI receives, as published by ariestar/sivtr in skills/sivtr-memory/SKILL.md and read by ahel’s review.
Sivtr is the shared local work memory for this machine.
Use it before asking the user to paste logs, repeat decisions, or restate earlier work.
Core rule:
Search for evidence first. Expand only the smallest relevant context. Ask when memory is missing, ambiguous, stale, or permission is required.
MCP tools (preferred when available)
If the host exposes sivtr MCP tools, prefer them over shelling out to the CLI. Register with:
sivtr mcp install -y
| Tool | Use |
|---|---|
sivtr_search | Narrow search over terminal / agent / provider / desk:... |
sivtr_show | Expand a ref or @last / @name |
sivtr_zoom | Neighboring record context |
sivtr_filter | Narrow @last or a saved set |
sivtr_status | Hooks, providers, daemon, local workspace origins (ws), remotes, vars |
Fall back to CLI commands below when MCP is not configured.
When to Use
Use this skill whenever local work memory may contain useful evidence. Sivtr retrieves recent terminal activity, AI conversation history, prior decisions, validation output, debugging trails, and handoff context.
Use it before asking the user to repeat context when you need to know things like:
- what just happened in the terminal or an agent session
- what command was run, what output it produced, or whether validation passed
- what the user or a previous agent decided, rejected, or planned
- where earlier debugging, build, test, lint, deploy, or research work left off
- what evidence supports a recap, handoff, status update, or next step
- whether missing or truncated current-context output exists in local memory
Do not use Sivtr as truth by itself. Treat memory as evidence to retrieve, then verify current files or commands before making claims about current state.
Default Retrieval Workflow
Mental model to keep in mind:
- A WorkSet is
records + anchors:recordsare the materialized backing store;anchorsare the active positions moving throughsearch,filter,work,nav,zoom, andshow. - Treat
@lastand@nameas WorkSet variables. Use--save <name>orsivtr var set <name> <source>whenever a result will be reused or refined. - Use
filterto narrow a WorkSet,work partsto project record anchors into part anchors,navto move anchors deterministically, andzoomonly when you want neighboring record context. - In shell pipelines,
@means "read the WorkSet JSON from stdin". Do not pipe--refstext into@; either omit--refsin intermediate commands or use@last/@name.
- Convert the user's vague reference into a small query.
- Choose a source:
terminal,agent, or a provider (codex,claude,cursor,opencode,openclaw,grok,hermes,pi,qoder,qodercn,dsh,gemini,goose,qwen), a WorkRef selector, an origin-prefixed source such asdesk:terminal/docs:agent, or a WorkSet variable such as@last/@name[1,3]. - Search is BM25-primary: a plain-text positional
QUERY(no regex) ranks the whole source by relevance;-m/--matchis an optional case-insensitive regex that bounds the set first, and the QUERY (or--matchtext alone) then ranks it. Use--last/--sincefor time windows,-i/--infor field filters, and--kindfor part kinds (prompt,command,user,assistant,tool,tool_call,tool_result,skill,thinking,output,error).- Latest terminal error:
sivtr s terminal --status fail --latest 1 --save latest_failure --refs - Broader terminal error scan:
sivtr s terminal "error" --latest 20 --save error_hits --refsorsivtr s terminal -m "Error|error|failed|fatal|not found|External command failed" --latest 20 --save error_hits --refs - Mounted remote memory:
sivtr s desk:agent "decision" --latest 20 --save remote_hits --refs
- Latest terminal error:
- Save/refine WorkSet variables instead of re-running broad searches:
sivtr filter @error_hits -m "more specific terms" --save narrowed --refssivtr filter @last[1,3] -m "narrower terms" --save focused --refs
- Move anchors deliberately:
- Filter a saved set:
sivtr filter @focused --status failure --save failed_focused --refs - Drill into parts:
sivtr work parts @focused --kind tool_result --save output_parts --refs - Deterministically move from a part to its record or neighbor:
sivtr nav @output_parts[1] '<+1>1' --refs - Return to records:
sivtr work records @output_parts --save parent_records --refs
- Filter a saved set:
- Expand only useful parent records:
sivtr zoom @focused[1] -C 2 --save ctx --refssivtr show @ctx --fullsivtr show <ref> --full
- Answer with evidence, then verify current files or commands when the claim depends on present repository state.
- Ask the user only after local memory has been checked and still lacks the needed fact.
WorkSet Flow
search, filter, work records, work parts, nav, and zoom create WorkSets. Each run saves the result to @last. Add --save <name> where supported, or use sivtr var set <name> <source> to keep a named WorkSet variable.
Core semantics:
recordsare backing facts/materialized context.anchorsare the active selection and the thing that moves through pipes.searchsearchesWorkParts, then outputs anchors at the current input granularity.filterapplies the same WorkSet filters to any source or piped WorkSet.navmoves anchors deterministically. It does not implicitly expand children; child movement must specify>N.showrenders at anchor granularity: record anchors show records, part anchors show just that part.zoommaps any anchor to its parent record, then expands nearby records.
Source forms:
terminal,agent,codex,claude,cursor,opencode,openclaw,grok,hermes,pi,qoderterminal/<session>/<record>,<provider>/<session>/<turn>,<provider>/<session>/<turn>/p<part>, and selector variantsorigin:bodyfor another local workspace or named remote, for exampledesk:terminal/...,docs:codex/4all:<src>searches every origin as one corpus (all local workspaces plus the current workspace's remote mounts):all:agent "workflow 最佳实践"@last,@name,@name[1],@name[1,3],@name[1..5],@name[1..3,8]@reads a WorkSet from stdin in shell pipelines
Default content search covers dialogue turns, terminal output, tool results (execution errors), and thinking; tool-call payloads and skill text stay out as noise and remain reachable with --kind tool_call / --kind skill or -i all. Structure channels (tool / skill / thinking) appear as <:channel:…:> markers in reading mode; expand with show --full or the TUI r toggle when you need payloads.
Output behavior:
- Terminal stdout with no explicit format prints
full. - Piped stdout with no explicit format prints WorkSet JSON for the next command.
--jsonis--format workset.--refsis--format refs.--fullonshowis--format full.
Pipeline example (pipe WorkSet JSON; do not add --refs in intermediate steps):
sivtr s agent -m "panic|failed" --latest 20 --save failures \
| sivtr filter @ -m "cargo|test" --save test_failures \
| sivtr zoom @ -C 1 --save failure_ctx \
| sivtr show @ -f timeline
Anchor movement example:
sivtr s pi -m "git push|main -> main" --latest 10 --save push_hits --refs
sivtr work parts @push_hits --kind tool_result --save push_outputs --refs
sivtr filter @push_outputs -m "main -> main" --save exact_output --refs
sivtr show @exact_output --full
Anchor navigation example:
sivtr s pi -m "panic|failed" --latest 10 --save hits --refs
sivtr work parts @hits --kind output --save output_hits --refs
sivtr nav @output_hits[1] '<' --refs # part -> parent record
sivtr nav @output_hits[1] '<+1>1' --refs # parent record -> next record -> first child
sivtr nav @output_hits[1] '<[-2..+2]' --refs # parent record sibling window
sivtr nav @output_hits[1] '~' --refs # containing session records
Named WorkSet variable example:
sivtr s agent -m "decision|TODO|next step" --latest 20 --save hits --refs
sivtr filter @hits -m "workset|zoom|show" --save narrowed --refs
sivtr var list
sivtr zoom @narrowed[1] -C 2 --save ctx --refs
sivtr show @ctx --full
Non-Interactive Safety Rules
- Prefer non-interactive commands:
sivtr s ... --refs,sivtr s ... -f <timeline|compact|md>,sivtr show ... --full,sivtr show ... --json. Prefer MCP tools when the host exposes them. - Use WorkSet variables (
@last,@name) and@pipelines for chaining; save reusable intermediate sets with--save <name>orsivtr var set <name> <source>. - Bare
sivtr(no subcommand) opens the interactive workspace browser — human-only. Do not open TUI pickers (--pick, hotkey, baresivtr) unless the user explicitly wants interactive selection. - Do not run
sivtr clear, hotkey start/stop, shell init, share/remote mutations, or config mutation unless the user explicitly asks. - Avoid clipboard-oriented workflows in agent retrieval. Use refs from
searchand expand them withshow/zoom. - Avoid dumping huge histories into the model. Search narrowly first (default bound is
--latest 5when neither--latestnor--limitis set), then expand only the relevant records. - If
sivtris not installed or no session log exists, say so briefly and continue with normal tools. Do not invent memory results.
Load References as Needed
References are relative to this skill directory.
references/commands.md— command syntax, WorkSet handling, and token budget.references/patterns.md— common user intents mapped to retrieval steps.references/evidence.md— what counts as evidence and how to report it.
Read only the file needed for the current task.
Signals
- GitHub stars
- 267
- Forks
- 20
- Last commit
- Sep 2026
Advanced
- Catalog kind
- skill
- Gateway key
sivtr-memory- Source
- github.com/ariestar/sivtr