@memosift/mcp-server

MCP (Model Context Protocol) server for MemoSift — context compression tools for AI agents.

MemoSift compresses LLM conversation context through a 7-layer pipeline with 7 compression engines, preserving tool call integrity and critical facts while reducing token usage by 2-5x. Zero LLM calls by default.

Verified on Real Production Data

| Metric | Result | |---|---| | Compression | 2.91x (coding) / 5.10x (general) on 11 real sessions | | Fact retention | 90.4% / 466 out of 466 fidelity probes pass | | Tool call integrity | 100% across 5.5M tokens, 4,799 tool calls | | Cost | $0.00 — zero LLM calls in deterministic mode | | Latency | <200ms per compression call |

Quick Start

npx @memosift/mcp-server

Claude Desktop

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "memosift": {
      "command": "npx",
      "args": ["@memosift/mcp-server"]
    }
  }
}

Claude Code

Add to ~/.claude/settings.json:

{
  "mcpServers": {
    "memosift": {
      "command": "npx",
      "args": ["@memosift/mcp-server"]
    }
  }
}

Cursor / Windsurf / VS Code

Same pattern — add to your editor's MCP server configuration.

Tools

The server exposes 8 tools that agents can discover and invoke:

memosift_check_pressure

Check if your context window needs compression. Returns pressure level (NONE/LOW/MEDIUM/HIGH/CRITICAL) and a recommendation. Call this before compressing to avoid unnecessary work.

| Parameter | Type | Description | |-----------|------|-------------| | model | string | Model name (default: claude-sonnet-4-6) | | current_usage_tokens | integer | Tokens currently consumed | | message_count | integer | Alternative: estimate from message count |

memosift_compress

Compress conversation messages to reduce context window usage. Accepts messages in any supported format (OpenAI, Anthropic, LangChain, Google ADK, Vercel AI). Returns compressed messages in the same format plus metrics.

| Parameter | Type | Description | |-----------|------|-------------| | messages | object[] | Required. Messages in any framework format | | preset | enum | coding | research | support | data | general | | model | string | Model name for adaptive compression | | framework | enum | openai | anthropic | agent_sdk | adk | langchain | vercel_ai (auto-detected) | | task | string | Task description for relevance scoring | | session_id | string | Session ID for persistent state | | current_usage_tokens | integer | Token usage for adaptive pressure | | system | string | System prompt (Anthropic format) |

memosift_configure

Create or update a compression session with specific settings. Sessions support incremental compression — cached state makes subsequent calls faster.

| Parameter | Type | Description | |-----------|------|-------------| | session_id | string | Required. Session ID | | preset | enum | Domain preset | | model | string | Model name | | token_budget | integer | Hard token limit | | recent_turns | integer | Turns to protect (1-20) | | enable_summarization | boolean | Enable LLM summarization |

memosift_get_facts

Retrieve critical facts (file paths, errors, decisions) extracted during compression. Facts survive even when source messages are dropped — powered by the Anchor Ledger.

| Parameter | Type | Description | |-----------|------|-------------| | session_id | string | Required. Session ID | | category | enum | Filter: INTENT | FILES | DECISIONS | ERRORS | ACTIVE_CONTEXT | IDENTIFIERS | OUTCOMES | OPEN_ITEMS |

memosift_expand

Re-expand a previously compressed message to see its original content. Useful when an agent needs to re-read a file it already saw.

| Parameter | Type | Description | |-----------|------|-------------| | session_id | string | Required. Session ID | | original_index | integer | Required. Message index from compression report |

memosift_report

Get detailed compression metrics from the most recent compression, including per-layer breakdown, adaptive overrides, and individual decisions.

| Parameter | Type | Description | |-----------|------|-------------| | session_id | string | Required. Session ID | | detail_level | enum | summary | layers | decisions | full |

memosift_list_sessions

List all active compression sessions with metadata (preset, model, fact count, last compression time).

memosift_destroy

Destroy a session and free its memory.

| Parameter | Type | Description | |-----------|------|-------------| | session_id | string | Required. Session ID to destroy |

Configuration

Environment Variables

| Variable | Default | Description | |----------|---------|-------------| | MEMOSIFT_SESSION_TTL_MS | 3600000 (1h) | Session TTL in milliseconds | | MEMOSIFT_DEFAULT_PRESET | general | Default compression preset | | MEMOSIFT_MODEL | — | Default model for new sessions | | MEMOSIFT_LLM_PROVIDER | — | openai or anthropic for Engine D | | OPENAI_API_KEY | — | Required if LLM provider is openai | | ANTHROPIC_API_KEY | — | Required if LLM provider is anthropic |

Presets

| Preset | Optimized For | Compression | Quality | |--------|---------------|:-----------:|:-------:| | coding | Code editing — preserves errors, file paths, signatures | 2.91x | 90.4% fact retention | | research | Research — preserves citations, URLs, findings | 3.6x | 92.0% quality | | support | Customer support — keeps error traces, resolutions | 2.4x | High recent turn preservation | | data | Data analysis — preserves schemas, query results | 3.6x | Numeric value preservation | | general | General-purpose balanced compression | 5.10x | 89.8% fact retention |

How It Works

The MCP server wraps MemoSift's MemoSiftSession — a stateful compression session that manages the anchor ledger (fact extraction), cross-window dedup state, compression cache, and incremental state across multiple tool calls.

Each session_id maps to an independent session. Sessions expire after 1 hour of inactivity (configurable). The _default session is used when no session_id is specified.

By default, the server runs in deterministic-only mode — no LLM calls, no external dependencies. All 6 deterministic compression engines are active. Engine D (LLM summarization) is only enabled when an LLM provider is configured.

Adaptive Compression

When a model is specified, the server uses Layer 0 adaptive compression:

| Pressure | Context Remaining | Behavior | |----------|:-----------------:|----------| | NONE | >60% | Skips compression — zero overhead | | LOW | 40-60% | Light: dedup + verbatim only | | MEDIUM | 25-40% | Standard: adds pruning, structural, discourse | | HIGH | 10-25% | Aggressive: all engines, observation masking | | CRITICAL | <10% | Maximum: auto-enables Engine D if available |

Cost Savings

| Volume | Opus ($15/MTok) | Sonnet ($3/MTok) | GPT-4o ($2.50/MTok) | |---|---|---|---| | 1M tokens/month | $9.82 saved | $1.96 saved | $1.64 saved | | 100M tokens/month | $982 saved | $196 saved | $164 saved | | 1B tokens/month | $9,816/mo | $1,963/mo | $1,636/mo |

MemoSift itself costs $0.00 in deterministic mode.

Community

• Discord — Join the MemoSift community
  • #general — Discussion and announcements
  • #help — Get help with integration and configuration
  • #feature-request — Suggest new features and improvements
  • #show-what-you-built — Share your MemoSift-powered projects
• GitHub Issues — Bug reports and feature requests
• memosift.dev — Documentation

License

MIT