MCP Server

4 tools that give Claude Code, Cursor, and any MCP-compatible client access to MemoSift memory. Recall, explore, fetch, and compress — all over stdio.

SETUP

One-command install for Claude Code

Authenticates and wires MemoSift into your Claude Code settings automatically.

terminal

memosift login
memosift install-claude-code

Manual setup (Claude Desktop / Cursor)

Add the following to your MCP settings configuration:

mcp-settings.json

{
  "memosift": {
    "transport": "stdio",
    "command": "memosift",
    "args": ["serve"]
  }
}

Or run directly

Start the MCP server manually over stdio.

terminal

memosift serve

Requires: pip install memosift[claude-code] (adds mcp dependency)

TOOLS

memosift_recall

3 modes

Search your memory for relevant knowledge from this session or across all sessions.

PARAMETERS

Name	Type	Required	Description
`query`	`string`	optional	Natural language search query
`intent_version`	`integer`	optional	Recall by intent version number
`turn`	`integer`	optional	Snapshot at specific turn
`scope`	`string`	optional	"session" or "project" (default: "session")
`mode`	`string`	optional	"fast" or "deep" (default: "fast")
`limit`	`integer`	optional	Max results, 1-100 (default: 10)

INPUT

input.json

{
  "query": "auth middleware decisions",
  "scope": "project",
  "mode": "deep",
  "limit": 5
}

OUTPUT

output.txt

Found 3 result(s):
Session intent (v3): Implement authentication flow

1. [memory] Auth module uses JWT rotation with 24h expiry (score: 0.87)
   Entities: AuthService, JWT
   Artifact: auth.py (code_python, 4.2KB)

2. [memory] Login endpoint at /api/auth/login (score: 0.72)
   Entities: AuthService, LoginEndpoint

3. [proposition] JWT tokens stored in httpOnly cookies (score: 0.68)
   Source artifact: auth.py

NOTES

•Set exactly one of query, intent_version, or turn.
•scope="project" searches across all sessions in the project.
•mode="deep" enables entity graph expansion for richer results.

memosift_explore

graph walk

Explore connections from a known memory, artifact, or proposition. Walks entity graph, same-artifact, temporal, and intent connections. No search query needed.

PARAMETERS

Name	Type	Required	Description
`item_id`	`string`	required	ID of the item to explore from
`kind`	`string`	required	"memory", "artifact", or "proposition"
`limit`	`integer`	optional	Max results (default: 10)

INPUT

input.json

{
  "item_id": "mem_abc123",
  "kind": "memory",
  "limit": 5
}

OUTPUT

output.txt

Exploring from memory mem_abc123:

1. [memory] Login rate limiting set to 5 attempts/minute (via: entity:AuthService, score: 0.82)
2. [proposition] Password hashing uses bcrypt with cost factor 12 (via: artifact:auth.py, score: 0.75)
3. [memory] Session tokens expire after 24 hours (via: temporal:turn_14, score: 0.71)

NOTES

•Use item IDs returned by memosift_recall.
•Walks 4 axes: entity graph, same-artifact, temporal proximity, and intent connections.

memosift_fetch

artifact

Download full artifact content with metadata. Content inline for <500KB, download URL for larger.

PARAMETERS

Name	Type	Required	Description
`artifact_id`	`string`	required	ID of the artifact to fetch

INPUT

input.json

{
  "artifact_id": "art_7f2a3b"
}

OUTPUT

output.txt

Artifact: art_7f2a3b
Type: code_python | Size: 4.2KB | Tokens: ~1,050
Title: auth.py
Summary: JWT auth middleware with token rotation
Tool: read_file | Turn: 12

Content (4,200 bytes):
import jwt
from datetime import datetime, timedelta
...

NOTES

•Content is returned inline for artifacts under 500KB.
•Larger artifacts return a time-limited download URL instead.

memosift_compress

no LLM

Assemble a structured session context block from stored knowledge. No LLM call — pure SQL + template. Organized by intent version (newest first).

PARAMETERS

No parameters \u2014 uses the current session automatically.

INPUT

input.json

{}

OUTPUT

output.txt

[MemoSift Session Context]

Current goal (v3): Implement authentication flow
Memories:
- Auth module uses JWT rotation with 24h expiry (importance: 0.9, turn 12)
- Login endpoint at /api/auth/login (importance: 0.8, turn 14)
Artifacts:
- art_7f2a3b: auth.py (code_python, 4.2KB) — JWT auth middleware

Previous goal (v2): Set up database schema
Memories:
- PostgreSQL 16 with pgvector extension (importance: 0.9, turn 3)

Key entities: AuthService (15 mentions), JWT (8 mentions)
Relations: AuthService --[uses]--> JWT

Use memosift_recall to search for details. Use memosift_fetch to download artifacts.

NOTES

•Uses current session automatically — no parameters needed.
•Pure SQL + template assembly, executes in <50ms.
•Organized by intent version with newest first.