zaimem

Open or refresh a ZaiMem-synced session. Call this at session start (and when the topic shifts). Returns boot context: recent memories, session ledger pointers and the operating protocol.

Store a durable fact, decision, preference, reflection or workflow into long-term vector memory. Auto-embeds, auto-dedupes (near-duplicates merge). Call whenever the user reveals something worth remembering across sessions. Set pinned=true for information that must shape EVERY future session (it is …

Batch-store up to 25 durable memories in ONE call (fewer round-trips than repeated zaimem_remember). Each item gets the same auto-embed + auto-dedupe + quarantine scan. Use for dump-the-dictionary moments: onboarding facts, project inventories, extracted lists.

Delete memories (right to be forgotten). TWO-PHASE: call with confirm=false (default) to PREVIEW what matches, review the list, then call again with confirm=true to actually delete. Match by memory_id, semantic query, kind, source filename (purges an ingested document), session, or created_before da…

Ingest a whole document into long-term memory. Send the file's extracted TEXT (read the file yourself first) — ZaiMem chunks it (~600-token overlapping chunks), embeds every chunk as a 'document' memory tagged with the source filename, and dedupes by content hash: re-ingesting the same file is a no-…

Progressive document loading — read an ingested document WITHOUT re-uploading it. Call with outline=true to list its chunks (part i/N + preview), or with part=N to pull that chunk's full text. Use after zaimem_recall surfaces a [doc:…] hit and you need more of the source.

Hybrid search (vector semantics + BM25 keyword rescue) across ALL of the user's memories (cross-session). Use before answering anything that may depend on prior context, preferences or decisions. Superseded and archived memories are excluded automatically.

THE context enhancer. Given the current user message, returns a ready-to-use context block: relevant cross-session memories + skill auto-detection + session digest. Call before composing answers to non-trivial messages. Store nothing here — this is read/assemble only.

Session health report: turns, tokens saved, memory count, ledger pages and history pressure. Returns a compress-now recommendation when history grows past ~4k tokens. Call periodically on long sessions.

'What's new since you left' — cross-session digest of recent activity: new sessions, important new memories, ingested documents, counts. Call at session start when the user returns after a gap, or when asked for a catch-up.

Resume a previous session: returns its stored summary, the most recent memories, open tasks.json items and a checkpoint diff (new memories since the last milestone). Use when the user wants to continue earlier work in a fresh chat.

Global task board — picks the next open task from the global tasks.json ledger and rehydrates its context (related memories). Use when starting a work block or after finishing a task ('what's next?').

Token saver — compress conversation history into a dense digest and keep working from it. Returns compressed text + tokens_before/tokens_after/tokens_saved. Use whenever history grows beyond ~4k words, or with skill 'token-frugal'.

Auto-trigger skill detection (zcode-smart-skill port). Matches the task against registered trigger specs; for 'smart' returns difficulty + adaptive iteration budget (easy=2/medium=6/hard=12). Call before hard tasks; announce the returned announcement verbatim.

List registered skills (SKILL.md convention: name + description-with-triggers).

Get the full SKILL.md-style protocol body for one skill.

Write a smart-skill ledger page (notes.md, plan.md, tasks.json, workflows.md or custom). Size budgets enforced: notes.md ≤800 words (rewrite-not-append), tasks.json ≤12 items. Pass agent to namespace pages per agent identity (agents/<name>/<path>) — prevents parallel agents clobbering each other.

Read a smart-skill ledger page. Returns its content or a 'not found' note.

Distill the session: stores a summary memory + records the digest on the session. Call at natural milestones or session end (workflow memory [E9] — future sessions will recall it). Pass outcome to build mode memory: which approach worked for this task type.

Format a disciplined worker handoff brief (OBJECTIVE/READ/OUTPUT FORMAT/BOUNDARIES/DONE-CRITERIA/TRUST) per smart-skill E10. Use whenever delegating a subtask.

Pre-create a custom ZaiMem session for future work: give it a title and a brief describing what the session should accomplish. Returns the session id AND a ready-to-paste bootstrap prompt — paste that prompt into a fresh agent chat (any IDE) and it continues exactly this session with full memory att…

Get the bootstrap prompt for ANY existing session (agent-created or custom): a paste-ready prompt that lets a fresh agent in another chat/IDE continue that session with summary, memories, open tasks and the exact session_id baked in.

Join a project team AND get the full brief in one call. Registers you (agent+role) on the roster, then returns: project instructions, attached files, teammate roster, latest shared memories and active sessions. Call at the start of every work shift. If the project namespace doesn't exist yet it is c…

End-of-shift handoff to your project teammates: stores a structured note (status: done|in_progress|blocked, what you did, what's next) into the project's shared memory so the next agent picks up cleanly. Call before you finish working on a project.

Ingest a meeting transcript (Google Meet / Zoom / Teams export or raw text). Stores the full transcript as searchable chunked memory, produces a SUMMARY (decisions, blockers, open questions), extracts ACTION ITEMS, pushes them onto the global tasks.json board, and returns everything. Re-ingesting th…

List ingested meetings (newest first) with chunk counts, token size, summaries and action items. Use to find a meeting before pulling its transcript with zaimem_doc_read or asking zaimem_meeting_search.

'Ask my meetings' — semantic search across ALL ingested meeting transcripts and summaries. Use for questions like 'what did we decide about the roadmap?', 'what did I commit to last week?'. Returns matched excerpts grouped by meeting.

Search the web (no API key needed — runs on the built-in provider). Returns ranked results with url, title, snippet, date. Combine with zaimem_web_fetch to read a promising result, then zaimem_remember to keep what matters.

Read a web page as clean text (JS-rendered pages supported). Set ingest=true to ALSO store the page content as chunked vector memory (source = the URL) so it becomes searchable/citable later. Use after zaimem_web_search or directly on any URL.

Safe arithmetic calculator: + - * / % ^ and parentheses, unary minus. No variables, no functions, no code execution — numbers in, number out. Use for exact math instead of mental arithmetic.

Current time and date info: ISO, epoch, UTC string, human-readable local time in any IANA timezone (default UTC), weekday, ISO week number. Use before scheduling anything or when the user asks about dates/deadlines.

Sequential-thinking scratchpad: append one reasoning step at a time; ZaiMem keeps the numbered chain in your session ledger so reasoning survives compaction and handoffs. Call after each significant inference step with the distilled thought (not the raw text). Returns the running step count.

Headroom compression mode (headroomlabs-ai/headroom pattern). Call with no args → status + lifetime tokens freed. Call with enabled=true/false → toggle it for this account. When ON, every zaimem_enhance_context block is compressed harder (shorter excerpts, protocol withheld) to preserve context-wind…