Session-start context: call this FIRST in a new conversation. Returns the resolved project scope with recent session snapshots, core memories, recent decisions, and open action items as a bounded digest.

List memory files, optionally filtered by category, core status or frontmatter. Use to browse what memories exist before reading or searching.

Read the full contents of a memory file. Use after palinode_list or palinode_search to see the complete content of a specific file.

Search Palinode memory for relevant context about people, projects, decisions, insights, or research. Returns the most relevant memory file excerpts ranked by configured hybrid or lexical retrieval.

Save a memory (fact, decision, insight, project update) worth keeping across sessions. Requires exactly one of type or ps=true. On timeout the save may still have committed — palinode_search a distinctive phrase before retrying, or you'll duplicate it.

Fetch a URL and save it as a research reference in Palinode memory.

Check Palinode health plus the effective read-only capture/recall pause state, policy provenance, and resolved project scope.

Show the change history of a memory file. Tracks renames (--follow) and includes diff stats per commit. Use detail='full' for the commit-level evolution view.

List all known entities, or get memory files referencing a specific entity.

Run a manual knowledge consolidation pass. Set dry_run=true to preview the proposed operations without applying them. A pass that reaches the LLM can run for minutes and may outlast your client's own tool-call timeout; the server finishes it either way and holds a run lock while it does, so a retry returns 409 rather than starting again.

Archive ephemeral memories whose expires_at has passed (ADR-015 §2.3 TTL regime). Deterministic + idempotent — flips expired memories to status: archived so they drop out of default recall while staying on disk. Set dry_run=true to preview.

Retire one specific memory that is wrong or obsolete. Sets status: archived so it leaves default recall, records the reason in the file's history sibling, and commits — never hard-deletes, so the content stays auditable. Pass superseded_by to name the memory that replaces it (a SUPERSEDE rather than a plain archive). Use this instead of re-saving a memory with a hand-written tombstone body: that leaves the wrong content live in search.

Bring one archived memory back into default recall — the inverse of palinode_archive for every archive path (on-demand, forget request, TTL expiry, consolidation). Flips status back to active, drops superseded_by, records restored_at / restored_from provenance and a history line, and commits. Does not un-strike retraction markers (use palinode_unretract) or re-enable triggers.

Withdraw one preference's mention-level retraction from one memory: un-strikes every ~~…~~ [RETRACTED …] span the pref produced and removes the pref from the file's retracted_prefs record, so a later forget request for the same pref can strike again. Pass the pref phrase as recorded in the file's history sibling. The file's status is never changed.

Take a forget request back. Given the memory that holds the request ('please forget that I…'), restores every memory it archived, un-strikes every mention it retracted, and archives the request record(s) so they stop acting as the retraction. Each step is its own audited commit; failures are reported per target.

Show what memories changed recently. Use to review what was learned, decisions made, or facts updated in the last N days.

Trace a fact back to when it was first recorded. Shows which session or commit created each line in a memory file.

Compose the full provenance lineage of a memory file into one view: source citations, when it was first saved and last changed, the supersession trail, typed contradiction/evidence links, and how often it has been recalled. Rows whose provenance is not yet captured render an honest placeholder. Read-only.

Explain one delivery of context: given the bundle_id a search receipt returned, show which memories were supplied, the exact revision of each (and whether its source changed since), the resolved scope, the calling surface, each record's disposition, and the coverage qualifiers. Fields that were never recorded are reported as unavailable with the reason, never guessed. This is supplied context only — no evidence that anyone acted on it is recorded. Read-only.

Revert a memory file to a previous version. Safe: creates a new commit preserving the old version in history. Defaults to dry run. A rollback that would undo a retirement (archive/supersede/retraction) is named in the preview and refused unless undo_retirements=true; prefer palinode_restore to bring a retired memory back.

Sync memory changes to GitHub for backup and cross-machine access.

Register or manage a prospective trigger for Palinode. When a future user message semantically matches the description, the specified memory file will be automatically injected.

Call at the end of a coding or chat session to capture key outcomes to persistent memory. Writes a session summary to today's daily notes and appends status to relevant project files. Provide a brief summary of what was accomplished, decisions made, and any blockers.

Scan memory for health issues: orphaned files, stale active files (>90 days), missing frontmatter fields, and potential contradictions. Returns a report without modifying files.

Advisory project-memory review. Composes the deterministic health signals (stale files, long-unresolved open questions, open contradictions, orphans, missing descriptions, wiki drift) scoped to a project, and proposes corrective ops (PROPOSE_ARCHIVE/UPDATE/SUPERSEDE). Read-only — proposes, never applies. Omit project to review the whole store.

List correction candidates mined from harness session transcripts: moments the user overturned a decision, rejected an approach and said why, or asked for something to be remembered. Each candidate quotes a bounded span of the user's own words with the session, turn and project it came from. Advisory and read-only — candidates are proposals awaiting review, never applied, and the list is empty unless the store's operator enabled transcript capture and named the transcript paths in config. Running a fresh detection pass is deliberately an operator action on the CLI or REST API, not something this tool can trigger.

Show exactly what correcting or retiring one memory would change — and change nothing. Returns the old text, the proposed new text, the affected document (and claim), its exact source revision, the rationale, where the correction came from, the project scope, the supersedes/superseded_by relation that would be recorded, every other record that quotes or derives from the target (reported, never rewritten), and the recovery command. Call this before palinode_correction_apply: the revision it returns is what apply requires back, so a target that changed in between is refused rather than silently merged.

Apply a correction previewed by palinode_correction_preview. Requires confirm=true and the expect_revision that preview returned; a target that changed since the preview, or a ref matching more than one memory, is refused with what it found rather than resolved by guesswork. Writes only through the existing validated path — the replacement is saved and the original is archived with superseded_by, so the original stays on disk, in git and retrievable as history. An ordinary later observation must NOT be routed here: this is the explicit, confirmed path, and it is the only one that retires anything. applied: "partial" means the replacement was saved and the original was NOT retired — do not re-run the correction; run the complete_command it returns.

Record that a correction candidate was reviewed and declined. Writes no memory: the candidate's queue row is marked dismissed, with the reason, and kept — which is what stops a later transcript scan proposing the same span again. A reason is required, because a dismissal with none is indistinguishable from a candidate nobody ever looked at.

Preview (default) or apply the undo of a correction. Without confirm this reads: it states what would be restored, what would NOT be deleted, and what cannot be reached at all — restoring a previous assertion, deleting history and undoing an agent's external actions are three different things and only the first is on offer. With confirm=true and the preview's expect_revision it brings the archived record back to status: active. It refuses to resurrect a record that was separately retracted or withdrawn by a forget request.

Given draft memory content the LLM is about to save, return the top-K existing memory files whose embeddings are semantically near it. Use BEFORE writing a new memory to decide 'create new' vs 'update existing'. Each result includes a strong_dup flag — when true (similarity ≥ 0.90), the existing file is a near-paraphrase and the LLM should usually update rather than create. Preprocessing strips wikilink syntax and the auto-generated ## See also footer so notes linking the same entities don't false-positive as duplicates.

Given a [[wikilink]] whose target file does not exist, return existing memory files semantically near the link target text. Use during wiki-maintenance passes to either propose a redirect (rename the link to point at an existing file) or to create the missing target file with informed context about its semantic neighbours. Accepts either [[name]] or bare name.

Given a memory file path, find the top-K semantically related files that are NOT currently linked to or from it (no existing [[wikilink]] in either direction). Use during wiki-maintenance passes to surface implicit relationships that no wikilink yet captures — the LLM can then propose new cross-links. Preprocessing strips wikilink syntax and the auto-generated ## See also footer so notes linking the same entities don't false-positive as related.

Given a topic phrase (not a file), check whether any wiki page already covers it. Returns {covered: bool, best_match: str | null, similarity: float}. Use BEFORE ingesting new content to ask 'is this already covered?'. Different framing from palinode_dedup_suggest: takes a short topic phrase rather than full draft content, and answers the binary 'already covered?' question.

Ask what memory holds RIGHT NOW about a question, or about one record. Returns the assertions that stand (with their evidence and source revisions), what replaced what, conflicts with every side intact, and what is explicitly unknown. Use instead of palinode_search when you want the current answer rather than a list of hits. Read-only; a tight budget can shrink the answer but never turns a conflict into a settled one. Retired records are left out, as in search, unless include_retired=true.

Fast palinode health check (<500ms). Skips network probes and canary writes. Checks path integrity, config consistency, and env-var drift. Use this first; call palinode_doctor_deep when results are unclear.

Full palinode health check including network probes and canary write tests. Takes 10-15s. Use when palinode_doctor reports unclear results or you need to verify the API, watcher, and service connectivity.

List, read, or activate versioned LLM prompts stored as memory files in the prompts/ directory. Use 'list' to browse available prompts, 'read' to view a specific prompt's content, or 'activate' to set a prompt version as active (deactivates others of the same task).

Return the dependency tree for a milestone or task slug, or list all unblocked items. Reads depends_on / blocks / parallel_with frontmatter from ProjectSnapshot files. Set unblocked=true to answer 'what can I work on right now?' across all slugs.