chat-recall

Stop syncing a project path. Nothing under it leaves the machine from the next sync on. Use it when the user says "stop syncing this repo", "don't upload anything from ~/work/client", "keep this project out of chat-recall". WHAT IT DOES: adds the path to the exclusion rules — on this machine, and on the account so every device the user syncs from honours it. Sessions ALREADY uploaded from that path stay until they are deleted; say so, and offer recall_forget for the ones that matter. There is deliberately no tool to REMOVE an exclusion: an agent may narrow what is synced, never widen it. Un…

Delete one conversation from chat-recall, everywhere, permanently. Use it when the user says "forget this conversation", "delete that session", "remove what we discussed about X" — anything that asks for a conversation to stop existing on the server. WHAT IT DOES: purges the session on every server this machine is logged in to, and tombstones the id so no later sync can bring it back. The AI tool's own transcript file on the user's disk is NOT touched — tell them that, because it is usually what they mean to keep. IRREVERSIBLE. There is no undo and no tool that restores it. Only pass `confi…

Read a key/value pair previously written with recall_set — or, when `key` is omitted, LIST the recorded keys (in one scope, or across all scopes when `scope` is also omitted). Use the list mode to discover what state has been recorded without remembering specific keys.

Report how many of your sessions are DAMAGED — the rendered conversation is thinner than the shrink-protected raw archive, which happens when an upstream tool truncates a transcript in place (a resume/compaction rewrite, or a script mirroring a shorter file over a longer one). This is the authoritative "0 remaining" check. Read-only by default; apply=true heals, and healing only ever makes a conversation FULLER.

Description is built at runtime.

Every improvement worth doing, ranked highest-priority first — the "what should I fix next, across everything" view. Merges the ranked code actions (recall_code_actions) with the non-instruction recommendations from account and project scope (focused reviews, project labels, POC resets). Code actions carry a numeric `pri` and recommendations carry a severity; both are mapped onto one scale so a single ordered list is honest. Set `create_tasks: true` to open one task per improvement on the shared team board (recall_tasks). It is off by default — this reads unless you ask it to write.

Index all Claude Code sessions for semantic search. Scans ~/.claude/projects/ and creates embeddings for all sessions. By default, only indexes new or changed sessions.

Add a fact to the knowledge graph. Subject → predicate → object with optional time window. E.g. ("example-app", "uses", "Postgres", valid_from="2026-01-15")

Mark a fact as no longer true. Use when things change — tools replaced, decisions reversed, people leave.

Query the knowledge graph for an entity's relationships. Returns typed facts with temporal validity. E.g. "example-app" → uses TypeScript, deploys to Kubernetes. Filter by date with as_of to see what was true at a specific point in time. Use this to VERIFY facts before asserting them.

Knowledge graph overview: entities, triples, current vs expired facts, relationship types.

Get chronological timeline of facts. Shows the story of an entity (or everything) in order.

Tag every user prompt in a session with sentiment / corrective markers: • interrupt — user hit ESC mid-response • frustrated — caps, profanity, "wtf", "ffs" • correction — "no", "stop", "don't" — negating the last action • approval — "yes", "do it", "i approve" • question — "why", "what", ends with ? • directive — "use X", "implement Y", "build Z" • clarification_request — "wdym", "explain", "what is" Returns per-prompt markers + a session-level summary count. Use this to spot sessions where things w…

Fetch ONE memory item, its full text, or its links — or browse a whole source type. Complements recall_memory_search: search finds candidates, this reads the specific item. mode=browse lists a type (e.g. every skill, every plan) without a query.

Search across all memory types: sessions, plans, tasks, CLAUDE.md files, history, paste cache, and agent diaries. Returns results from any memory source, ranked by relevance. Use source_types to filter. Plans and tasks are found here (source_types: ['plan', 'task']); read a full plan with recall_show (pass the plan id).

Aggregate session outcomes (shipped / interrupted / abandoned / in_progress) over the last N days — "how much of my work actually landed?".

Get the full project dossier as markdown — rich project context. Aggregates everything the index knows about one project into a single report: overview (from CLAUDE.md), tech stack (KG uses), architecture / deployment / security sections (from CLAUDE.md), decisions log (KG chose/rejected), recent session activity with summaries, open tasks, plans, agent diary conclusions, and cost rollup. Accepts EITHER a project path / name substring (e.g. "api", "example-app", "/home/user/code/example") OR a stable project_id: - "git:<host>/<owner>/<repo>" (e.g. "git:github.com/owner/repo") - "ws:<na…

Label a project POC, production or engineering — the guardrails every future session reads. This is the highest-leverage single call in the surface: the label tells any assistant working here whether it may reset a database and move fast (poc), must protect data and avoid destructive steps (production), or should hold a higher bar on tests and review (engineering). ASK THE USER before setting 'production' or clearing a label. Both change what other agents will consider permitted, and that is not yours to decide alone.

List recent sessions across every AI tool chat-recall indexes. Shows your most recent conversations across all projects or filtered by project. Use this when the user says "continue our last conversation" or wants to see recent work. Pass `since_hours: N` to restrict to sessions modified in the last N hours.

Re-run the memory classifier over already-indexed chunks so classifier improvements reach old data. Idempotent — only rewrites tags that actually changed.

Apply a recommendation: add its rule to the repo's CLAUDE.md, or set the label it asks for. Use it when a recommendation from recall_recommendations is right. You do not need to ask permission for a CLAUDE.md rule — it is additive, visible in the diff, and the user can delete a line. DO ask first for anything that changes what the AI is allowed to do destructively. A CLAUDE.md edit is queued for the machine that has the repo and lands within about 45 seconds; only that machine has the file. A label applies immediately. Applying twice is harmless.

Say no to a recommendation for this project, with a reason. Use it when advice does not apply to THIS repo — a reuse rule on a repo that is deliberately duplicated per environment, a label on a scratch project. Dismissed recommendations stop being offered for that project and no other. TELL THE USER what you dismissed and why. This changes how future sessions treat their codebase, and a reason nobody sees is a decision nobody can review. Pass undo: true to put one back.

Actionable recommendations, two scopes on one tool: - scope "account" (default) — from YOUR chat-recall data: leaked secrets found by the scanner + session behaviour (unresolved/abandoned). Returns concrete CLAUDE.md rules to apply globally. - scope "project" (requires `project`) — behavior × code recommendations for a code-indexed project. Concrete changes to apply: add a CLAUDE.md rule, install a skill, set a project label, or reset a POC db. Reasons over BOTH the code findings and how sessions in this project went (failed/abandoned, recurring corrections). Each has rationale + …

Before creating a file, check whether you have written one like it before. Scores the filename against every file touched in your synced history. Answers "have I built this already?" — the reinvention check.

Force a fresh AI summary for one session — use when its stored summary is missing, stale, or plainly wrong.

Give a conversation a memorable name — or rename it — so you (and the user) can find and resume it later. Mirrors Claude Code's session naming (`/rename`). The name replaces the auto-generated summary in recall_recent and the web UI. Use it when the user says "name this conversation X", "call this session Y", or "rename this to Z". Pass an empty string to clear the name and revert to the auto title. The name persists across re-syncs and summary regeneration (it's stored separately from the AI summary).

Search for relevant past sessions to resume. Find past conversations matching your current task. Returns session IDs that can be used to resume that session in the tool that produced it. Three modes on one tool: - `query` — free-text search ("I'm working on X, what past work is relevant?"). - `like_session` — find sessions similar to an existing session id (uses its first prompt as the search text, excludes itself, groups hits by project). - `include_outcome: true` — append a per-result outcome one-liner (shipped / interrupted / abandoned + edit stats) to judge whether each hit is worth …

Mark a secret finding as rotated, false_positive, or dismissed. Use when the user confirms a key was rotated or that the match is not a real secret. The dismissal syncs across devices.

List tenant custom secret-detection rules, or test a regex in the sandbox. Tenant rules are configured in the chat-recall dashboard; this tool exposes them to the agent and lets you test new regexes without persisting them.

Get secret findings for a specific session. Use after recall_search or recall_security_summary to show exactly which lines in a session matched which detector.

Get a security overview from the synced chat-recall server. Returns the action-required list of distinct leaked secrets (grouped by redacted preview), plus per-detector totals and top rules. Use this when the user asks "do we have any leaked secrets?" or "security findings" or "what should I rotate". Call this first; then call recall_security_session for specific sessions.

Persist a small key/value pair across sessions. Use this for state that doesn't fit the diary (narrative) or knowledge graph (entity-relationship facts) — things like "current PR url", "branch I'm working on", "user's preferred test runner". Scope namespaces keys so per-project state doesn't collide.

Which projects are shared into your team. scope=mine for your own shares, scope=all for every share visible to you. Private by default — nothing is visible to teammates until shared.

Get conversation content from a specific session — or the full text of a plan. Use this after recall_search to get full context from a session. Also accepts a plan id (the plan filename without .md, as returned by recall_memory_search(source_types:['plan'])) and renders the complete plan text. Four ways to reach a message: 1. `query` finds the messages that contain a text and gives their lines. 2. `around_line` returns max_messages messages centred on a line. 3. `from_end: N` returns the last N messages. 4. `expand_line` returns one message whole. Message text and commands come back in full.…

Get structured resume context for recent work — by default the last 3 sessions. Call it with NO arguments to resume. That is the common case: "continue", "pick up where we left off". Pass session_id to resume one specific session, or project_filter to scope to one project. Returns, for the newest session: - What was done (completed work, decisions made) - What's pending (unfinished tasks, TODOs mentioned) - Files modified with change summary - Token/cost budget used - Resume command Then a digest of each earlier session: its project, outcome, summary, the decisions it took and the task list…

Show index status and statistics: synced sessions, chunks, freshness, top projects, plus the full memory-system breakdown (items/chunks per source type, per AI tool, link count).

Search inside SUBAGENT transcripts — the explore/compact/aside sub-tasks that run within a session and whose work never appears in the main conversation. Use when a session's own turns don't contain what you remember: the finding may have come from a subagent. Indexed server-side, so it works across devices.

Get the summary for a session — including its OUTCOME classification. By default returns a *rich* summary that combines the AI summary (when available) with the structured outcome derived from the transcript itself: status (shipped / interrupted / abandoned / in_progress), the session window, decisions the agent announced, blockers hit (tool errors, interrupts), edit/commit stats, prompt markers, and the last assistant claim paired with the user's reaction to it (so you can see whether "done!" was actually accepted or met with "wtf"). This is the triage call when scanning a session list — it …

Leave a progress note on a team task — a comment visible to everyone on the board. Use it to record what you did, what is blocked, or why a status changed, without editing the task itself. Pass the task id (t_…) from recall_tasks.

Create a collaborative team task. Optionally assign it to a teammate and attach it to a project. Returns the new task id. When you then START working on a task, use recall_task_update to set status 'in_progress' and link your current session (linked_session_id) — that is what lets the board verify the work actually shipped.

Read or change what the board FILES — call with no arguments to read it. max_pri severity floor, inclusive: 0 critical, 1 high, 2 medium, 3 low ceiling how many auto-filed cards may be open at once max_per_run cap on new cards per run categories only file these categories (e.g. ["security"]); [] means all project apply max_pri/categories to THAT REPOSITORY ONLY — one floor across every repo forces the strictest one on all of them exclude_project / include_project a repository that never files a card — a client's codebase, an …

Update a team task: change status, reassign, rename, link a session, and/or add a comment. Pass the task id (t_…) from recall_tasks. WORKFLOW — do this whenever the user asks you to work on a task from the board: 1. When you start, set status 'in_progress' AND pass your own current session id as linked_session_id (the session id you know from your own context; this MCP server cannot see it). This is REQUIRED — claiming without it is refused, because "in progress" with nothing behind it cannot be asked about. The card then shows that session's changes as they land, not only after t…

List the task board — and, with detail, the work itself. Without `detail`: one line per card (title, id, assignee). Use for "what's on the team's plate", "my tasks", or to find a task id to update. With `detail: true`: the full brief per open card — the fix, the file locations and the agent prompt that the auto-filer already wrote — plus how to claim a card and how it gets closed. Use that whenever you intend to DO the work rather than report on it. Closing is earned: 'done' is refused without a linked session. Rejecting is the user's call, never yours.

See what each teammate did, grouped by project — the team activity view. Returns a per-member × per-project rollup (session counts + last activity) for your team. Scoped to what you're allowed to see: your own work plus teammates' work on projects they've shared into the team (private projects never appear). Use for "what has the team been working on", "what did <teammate> do on <project>", or standup/status. Filter by project, by member, or by recency (since_days).

Which MCP servers, skills, commands and agents exist, on which AI tool, on which device. Use it before recall_toolkit_sync to see what is missing where, and after, to confirm what landed. Counts are per (type, tool) and rows name the device that uploaded them, so you can tell "my laptop has it" from "my desktop has it". NOTE: the server holds an INVENTORY. MCP registrations carry enough to rebuild elsewhere; skills and agents do not (their file content is not uploaded), and this tool says which is which.

Install the account's MCP servers / skills / agents onto a device, or fan a device's own artifacts out across its tools. TWO DIFFERENT OPERATIONS — pick with `scope`: scope: 'pull' Install what the ACCOUNT has onto the target device, sourced from the server. This is the cross-device one: it is what sets up a new laptop. Omit `device` to target every device. scope: 'fan_out' Copy what ONE device already has between that device's own tools (claude → codex → cursor …). Single machine only; it reads that ma…

List the human-typed prompts from a session (or recent sessions), newest first. Tool results and system banners are stripped — you get only what the user actually wrote. Useful for "what was I asking yesterday?" or "what did I tell the agent in this session?". Pass `query` for "did I say X?": it keeps only the prompts that contain that text. Each prompt prints whole up to 2000 characters; a longer one prints its start and end and the recall_show expand_line call for the rest.

Build startup context for an AI session: identity blurb + the top classifier-flagged facts (decisions/preferences/milestones at importance >=4) + current knowledge-graph facts. Replaces a manual context dump. Call this once at the start of a session to "remember who you are and what's true right now".

Get a weekly activity digest across all projects. Returns: - Session count, total cost, coding time - Top projects by activity - Cost trend vs previous week - Open tasks across projects - Git activity summary Use to understand overall productivity and spending.