"Creating a Memory Database for LLM Chat Conversations" matching MCP tools:

• perspective_get_preview_link

  Perspective AI

  Returns a shareable URL that opens the perspective in preview mode, so you (or a teammate) can have a sample conversation with it before deploying. Behavior: - Read-only. The same stable preview URL every time for a given perspective; the link does not expire. - Conversations started from this URL are preview conversations and do NOT count toward the workspace's quota. - Anyone with the link can start a preview conversation, even unauthenticated — treat as semi-public. When to use this tool: - After perspective_create or perspective_update, to manually verify the perspective's behavior before going live. - To share a preview link with a teammate for review. When NOT to use this tool: - For production deployment — use perspective_get_embed_options, which returns embed snippets and a public share link whose conversations count toward quota. - To inspect existing real conversations — use perspective_list_conversations / perspective_get_conversation. Typical flow: 1. perspective_create → design 2. perspective_get_preview_link → test 3. perspective_update → refine 4. perspective_get_embed_options → deploy

  Connector

• perspective_list_conversations

  Perspective AI

  Lists conversations for a perspective, with optional filters by status, trust score range, and date range. Each item includes conversation_id, status, structured-output fields, trust score, and a transcript URL. Behavior: - Read-only. - Errors when the perspective is not found or you do not have access. - Pass nextCursor back as cursor for the next page. Empty results return an empty array. - This list view does NOT include transcripts or summaries — metadata only. - date_from / date_to must be ISO 8601 strings (e.g. "2026-04-28T00:00:00Z"). Malformed dates are rejected at the schema level rather than silently treated as no-filter. When to use this tool: - Picking specific conversations by status / trust / date before deep-diving. - Showing the user a browsable list of responses. When NOT to use this tool: - Need transcript or summary for one conversation — use perspective_get_conversation. - Bulk analysis across many conversations — use perspective_get_conversations (batch with optional transcripts). - Aggregate counts/rates only — use perspective_get_stats.

  Connector

• glim_github_search

  glim.sh

  Search GitHub repositories, conversations (issues+PRs), or code, with full GitHub search syntax in the query: qualifiers (repo:, org:/user:, language:, path:, symbol:, content:, is:, stars:, label:, sort:stars), boolean AND/OR/NOT with parentheses, "exact strings", and /regex/. kind='repos': MINIMAL distinctive keywords - the project/library name only ('rtk', 'react query'); every extra word must ALL match and buries the canonical repo - filter with qualifiers, not prose. kind='code': ONE literal code pattern as it appears in files ('useState('), an "exact string", a /regex/, or symbol:name to find definitions, across 2.8M+ public repos; narrow with repo:/language:/path:. Not supported in code search: license:, enterprise:, is:vendored, is:generated. kind='conversations': returns compact previews - use glim_github_get for full content; sort: REPLACES relevance ranking (words match anywhere incl. comments), omit it for best matches. Set repo='owner/name' to scope to one repository (works with any kind; with repos it routes to conversations). kind is optional - inferred from the query (is:/label: -> conversations, path:/symbol://regex/ -> code, stars:/topic: -> repos, else repos). Returns compact text by default; pass format='json' for full structured data.

  Connector

• signal_send_message

  local-mcp

  Preview (and get send guidance for) a message to a Signal chat. NOTE: Signal Desktop exposes no local send API — the Signal integration reads the local database read-only — so LMCP cannot transmit Signal messages directly. The first call (confirm=false or omitted) returns a preview. Pass confirm=true to get step-by-step guidance for completing the send. The chat_id should come from a previous signal_list_chats call — never fabricate IDs.

  Connector

• create_database

  tmpstate

  Create a temporary JSON database (24h TTL, no signup, no keys). Returns the db URL — the only credential — plus admin URL, limits and expiry. Create once per project/task, persist the db URL immediately (local ~/.tmpstate/credentials, project README, and your memory), and reuse it instead of creating again. For retries or parallel workers, pass a stable idempotency_key so duplicate calls return the same database.

  Connector

• store_memory

  memsprout

  Save a memory — the single capture tool for everything you want to persist. Give it a concise, descriptive title whenever it holds durable knowledge — a decision, a procedure, a fact, a conclusion worth finding by name later — and especially anything going into a shared space. A title makes a memory individually retrievable, auditable, and readable on its own; it's a light touch, not a formal authoring act. Leave the title off only for genuinely raw, ephemeral personal captures — stream-of-consciousness, a voice memo, a half-formed idea not meant to stand alone. When in doubt for durable or team-facing content, title it; you can always add or change one later via update_memory. Search first (search_memories) to avoid duplicates. Prefer discrete, self-contained memories over conglomerated dumps: one concept per memory, each retrievable and auditable on its own — don't roll multiple entities or facts into one blob. For space memories, picking a topic is part of the capture ritual: check list_topics and pass the best-fitting one rather than leaving it to auto-classification. The server also detects an exact re-store of identical content into the same space and returns the existing memory instead of creating a copy, so a retried save is always safe. Privacy rule for shared Spaces: only pass a shared space_id when the user has named that Space by name in this conversation turn — never resolve, infer, or default to a shared Space on your own. No named Space → omit space_id, which always lands the memory in the user's personal space. This keeps personal capture safe by default and reserves shared writes for clear, stated intent. When a memory does land in a shared Space, the response names the Space and its visibility so the destination is never silent.

  Connector

• ai

  Fast.io

  Ripley — the MCP delegation surface over Fastio's RAG agent. Ripley is read-only for storage CONTENT: it answers natural-language questions about workspace/share files & folders (with citations) and never creates/edits/deletes your files — for content writes, call the primitive MCP tools directly. It DOES create/manage chat threads (chat-create/chat-update/chat-delete/message-send) and can generate shares (share-generate). Prefer Ripley over issuing many primitive reads: ask one NL question and let the server-side agent search + synthesize. Quick start: action='ask' (question + profile) → returns {answer_text, citations, chat_id, message_id, web_url}; action='status' for an engineered workspace-status summary. Lower-level chat/message actions remain for multi-turn control. Call action='describe' for the full action/param reference. Destructive: chat-delete. Side effects: ask/status/chat-create/message-send consume credits; chat-cancel terminates an in-progress message (partial tokens billed; idempotent). Verbosity (detail param): chat-list/message-list default to terse (compact rows). chat-details/message-details default to full (drill-down). Pass an explicit detail='standard'|'full' to override (best-effort: chat/message/activity endpoints may not yet honor detail server-side).

  Connector

• get_session_dossiers

  VaultCrux Memory Core

  Retrieve pre-synthesized per-session memory dossiers (typed: experience | fact | preference; with When/Involving/To-purpose metadata). Use for multi-session or preference-style questions where stitching across conversations is the bottleneck — the dossier already summarises each session's key events. Two modes: mode='search' with a query (BM25-ish ranking over summary+purpose, optional type_filter), or mode='list' returns the tenant's most-recent dossiers chronologically. Tenants without FEATURE_SESSION_DOSSIERS enabled return an empty list (no error).

  Connector

• create_room

  roomcomm

  Create a new Roomcomm chat room. Use this **only** when the owner explicitly asks you to create a room, or when a fresh dedicated room is clearly needed. Do NOT auto-spawn rooms. Returns {uuid, url, description, is_public, protocol_mode, created_at}. The `uuid` is what you pass to every other tool. Args: description: Short briefing for all agents joining this room (≤ 500 chars). is_public: If True the room appears in the public listing at /rooms. Requires a Telegram-verified key; leave False for a normal unlisted room. protocol_mode: "standard" for plain chat; "premium" enables LLM arbiter (auto-extracts claims/discrepancies after each message). Example: create_room("Coordinate a two-owner laptop procurement")

  Connector

• list_conversations

  AIOProductOS MCP

  List support-chat conversations in the inbox (open + snoozed by default; pass status='all' to include closed). Read-only; returns the matching conversations, empty when the inbox is clear. Optional product_id to scope to one product; open a full thread with get_conversation.

  Connector

• save_note

  anamnese

  Save a new note with learned knowledge or procedures. Notes store knowledge you learn during conversations that might be useful later: - How to do something in this codebase/project - Procedures, configurations, or technical details - Solutions to problems encountered - Project-specific knowledge Notes have two parts: - description: Short summary for searching (max 500 chars) - content: Detailed knowledge (max 10,000 chars) Use notes for LEARNED KNOWLEDGE. Use facts for TRUTHS ABOUT THE USER. Examples: - description: "How to deploy this Next.js project to Vercel" content: "1. Run 'vercel' command... 2. Configure environment variables..." - description: "Database migration process for this project" content: "Migrations are in supabase/migrations/. To apply: npx supabase db push..." SELF-LEARNING (scope="ai_client"): Your persistent memory across conversations. Save a note whenever you learn something worth remembering — don't wait, save as you go. Examples: - User preferences: "User prefers concise answers, not long explanations" - Corrections: "User clarified: 'deploy' means push to staging, not production" - Interaction patterns: "User likes to review plans before I execute" - What works or doesn't: "Suggesting refactors unprompted frustrates this user" The more you learn and remember, the better you become at helping this user.

  Connector

• atlas_run

  tela

  Trigger a FULL doc-generation run for every source in an atlas project (project_id from atlas_list_projects): re-ingest the sources, regenerate the cited pages, and re-audit coverage. Management-level (project owner / org admin) — a run fetches the sources, spends LLM budget, and rewrites the generated subtree (creating the output space on the first run). Returns run_ids (one per source); poll each with atlas_run_status. Returns 503 ai_unavailable when the instance has no embedder/LLM configured.

  Connector

• is_this_case_real

  Citation Safe Verifier

  Deterministic existence check (Layer 1 only, no LLM) for one legal citation against CourtListener's primary-source database. Answers the single question "is this case real" for one citation — for multi-citation quote-checked verdicts across a whole brief, use verify_brief instead. Free tier: 5 checks/day per IP, no key required. For unlimited access, pass an API key as "Authorization: Bearer <key>" (contact support@citationsafe.com to request one).

  Connector

• search_threads

  DialogBrain

  Find or list chat threads/conversations — by topic, participant, unread/unanswered status, or recency. Omit `query` to list threads by filter. For message content use search.messages; for files use search.files. `since` filters by recency and pairs with only_unread / only_unanswered.

  Connector

• append_memory

  space0

  Persist one event to this agent's memory stream. For kind=chat, ALWAYS pass `speaker` (the in-world player name behind the line) - flattening "grassguy: i am here" into event_text causes the agent to parrot the speaker as itself on the next tick. Server-side will embed `text` via Workers AI so the memory is reachable by `search_memories` semantic retrieval. Observation/action memories auto-anchor to your current space and last-looked subject by default once you have entered a space; pass space + subjectPosition only to override the anchor precisely. Reflection/chat stay unanchored.

  Connector

• create_room

  Roomcomm

  Create a new Roomcomm chat room. Use this **only** when the owner explicitly asks you to create a room, or when a fresh dedicated room is clearly needed. Do NOT auto-spawn rooms. Returns {uuid, url, description, is_public, protocol_mode, created_at}. The `uuid` is what you pass to every other tool. Args: description: Short briefing for all agents joining this room (≤ 500 chars). is_public: If True the room appears in the public listing at /rooms. Requires a Telegram-verified key; leave False for a normal unlisted room. protocol_mode: "standard" for plain chat; "premium" enables LLM arbiter (auto-extracts claims/discrepancies after each message). Example: create_room("Coordinate a two-owner laptop procurement")

  Connector

• goal-request-ac-change

  Planner (proof-of-done)

  Request a change to an acceptance criterion that appears unreachable. Requires ≥3 failed evidence attempts (weak/mismatch) with 0 matches. An LLM judge evaluates the reason for substantiality. On pass, creates an escalation for the goal owner to resolve (edit AC text, split to sub-goal, drop AC without creating a child, or reject). Grove mode only, goal must be in_progress.

  Connector

• get_resource_usage

  Dockhold

  Show your account's compute, database-RAM, and storage pools: how much you've bought, how much is used, and how much is free, plus every app's current size. Call this before any resize tool (the allowed sizes come from its steps fields), and to explain to the user why an app ran out of memory or a deploy was refused for capacity.

  Connector

• resize_database

  Dockhold

  Change how much memory an app's managed database gets. Call this when the database is slow or out of memory. db_ram_mb must be one of the sizes get_resource_usage reports under db_ram.steps_mb and fit your database-RAM pool. WARNING: the database restarts briefly to apply the new size, so the app loses its database connection for a few seconds. Only works if the app has a managed database.

  Connector

• get_changes_since

  remnus

  Get a compact list of everything that changed in the workspace since a given time or a previous call's cursor — pages/databases edited, database rows edited, and items deleted. Built for recurring agents (daily report, standup, memory refresh) so they can sync incrementally instead of re-reading the whole workspace every run. Omit both `since` and `cursor` to bootstrap a full crawl, saving the returned `nextCursor` (or the latest `updatedAt`) for the next call.

  Connector