Webrenew Memories Memories MCP

memories-mcp

Connect AI agents to the memories.sh memory layer via MCP (Model Context Protocol).

The CLI is the primary interface for memories.sh — use memories generate to create native config files for each tool. The MCP server is a fallback for real-time access when static configs aren't enough. It's also the best choice for browser-based agents (v0, bolt.new, Lovable) where the CLI can't run.

Quick Start

# Local stdio transport (most reliable)
memories serve

# HTTP/SSE transport (for web clients like v0)
memories serve --sse --port 3030

# Cloud-hosted (no local install needed)
# Endpoint: https://memories.sh/api/mcp
# Header: Authorization: Bearer YOUR_KEY

Primary Tool: get_context

Always start with get_context — it returns active rules + relevant memories in one call:

get_context({ query: "authentication flow" })
→ ## Active Rules
→ - Always use TypeScript strict mode
→ ## Relevant to: "authentication flow"
→ 💡 DECISION (P) abc123: Chose JWT for stateless auth

Leave query empty to get just rules. Use limit to control memory count (default: 10).

For lifecycle-aware callers on local CLI MCP, pass compaction/session hints:

get_context({
  query: "checkout timeout",
  session_id: "sess_123",
  budget_tokens: 6000,
  turn_count: 6,
  turn_budget: 24,
  last_activity_at: "2026-02-26T23:00:00.000Z",
  inactivity_threshold_minutes: 45
})

These hints let the server trigger write-ahead checkpointing before destructive compaction.

When relationship extraction is enabled server-side, get_context may also return conflicts[] for contradiction-linked memories. Treat these as clarification prompts before taking irreversible actions.

Tool Selection Guide

Memory Types

When using add_memory, pick the right type:

• rule — Coding standards, preferences, constraints (always returned by get_context)
• decision — Architectural choices with rationale
• fact — Project-specific knowledge (API limits, env vars, etc.)
• note — General notes (default)
• skill — Reusable agent workflows (use with category and metadata)

Scopes

• project (default) — Scoped to current git repo, detected automatically
• global — Applies everywhere, set global: true in add_memory
• project override — Set project_id: "github.com/org/repo" in add_memory (or start_memory_stream) to force project scope when the MCP process is running outside that repo

Do not send both global: true and project_id in the same call.

Streaming Memory Tools

For collecting content from SSE sources (v0 artifacts, streaming responses):

1. start_memory_stream({ type?, tags?, global?, project_id? }) → returns stream_id
2. append_memory_chunk({ stream_id, chunk }) (repeat for each piece)
3. finalize_memory_stream({ stream_id }) → creates memory + triggers embedding
4. cancel_memory_stream({ stream_id }) → discard if aborted

Lifecycle Tools (Local CLI MCP)

These tools are currently available when running memories serve locally:

1. start_session({ title?, client?, user_id?, metadata?, global?, project_id? })
2. checkpoint_session({ session_id, content, role?, kind?, token_count?, turn_index?, is_meaningful? })
3. end_session({ session_id, status? })
4. snapshot_session({ session_id, source_trigger?, slug?, transcript_md?, message_count?, meaningful_only? })
5. consolidate_memories({ types?, include_global?, global_only?, project_id?, dry_run?, model? })

MCP Resources

For clients that support MCP resources:

Transport Options

Local-only tools: lifecycle + reminders + streaming. Hosted MCP focuses on tenant-routed core memory tools.

Reference Files

• Client setup configs: See references/setup.md for copy-paste configs for every supported client
• Full tool reference: See references/tools.md for all parameters, return formats, and examples

Converted and distributed by TomeVault — claim your Tome and manage your conversions.