Memory Forge

Memory Forge is a local-first MCP memory backend for MCP-capable AI clients. It gives compatible clients a small set of tools for saving, searching, and summarizing durable project memories without sending the database to a cloud service.

V1 is intentionally boring in the best way: a Python MCP server, SQLite, and full-text search.

Features

• Local SQLite database with FTS5 search.
• MCP tools for remember, search, context, update, and forget.
• Soft-delete by default so accidental forgets can be recovered from backups.
• Project, tag, source-agent, and importance metadata.
• Token-budgeted context retrieval to avoid duplicate long-term memory blocks.
• Usage estimates on retrieved context so clients can see read cost.
• Active-context compaction through memory_compact when clients send the working context they want Memory Forge to own.
• Model-window budgeting with context_window_tokens, reserved_prompt_tokens, and reserved_output_tokens.
• Fallback retrieval for focused queries that miss but project memory exists.
• Works with any MCP client that can launch a stdio server.

Memory Budgeting

Memory Forge returns only the focused memory a client asks for. Use small memory_context budgets by default, then request more only when the task needs broader project orientation. A good starting point is max_chars: 2000 for normal coding tasks and max_chars: 6000 for broad project orientation.

If the client knows the model window, pass context_window_tokens, reserved_prompt_tokens, and reserved_output_tokens so Memory Forge budgets retrieved memory against the same window as the model call.

See Memory Model Guide for token-budget guidance and active-context compaction behavior.

Install

uv sync

Run the MCP server:

uv run memory-forge

Configure Codex to use Memory Forge and disable duplicate built-in memory:

uv run memory-forge-setup codex --from-checkout .

For an installed release, run:

memory-forge-setup codex

Preview the config change without writing:

memory-forge-setup codex --dry-run

By default the database is stored at:

%USERPROFILE%\.memory-forge\memory.db

Set a custom path with:

$env:MEMORY_FORGE_DB="C:\path\to\memory.db"
uv run memory-forge

MCP Tools

memory_remember

Save a durable memory.

{
  "content": "The API service uses SQLite for local development.",
  "tags": ["api", "local-dev"],
  "project": "example-project",
  "source_agent": "codex",
  "importance": 3
}

memory_search

Search memories with optional filters.

{
  "query": "SQLite local development",
  "project": "example-project",
  "tags": ["api"],
  "limit": 10
}

memory_context

Return compact prompt-ready context for an agent.

{
  "query": "database setup",
  "project": "example-project",
  "limit": 8,
  "max_chars": 2000,
  "context_window_tokens": 128000,
  "reserved_prompt_tokens": 24000,
  "reserved_output_tokens": 4000
}

The response includes context, count, max_chars, truncated, and the raw matching memories. It also includes usage with character count and rough token estimates, including whether the returned memory fits the declared model window budget. If a focused query misses but project or tag filters can still return relevant memories, fallback_used is true. Clients should inject only the context string into the working prompt.

memory_compact

Compact active context supplied by a client and optionally store the compacted result in Memory Forge.

{
  "active_context": "Current chat notes or working context supplied by the client.",
  "project": "example-project",
  "tags": ["handoff"],
  "source_agent": "codex",
  "max_chars": 2000,
  "context_window_tokens": 128000,
  "reserved_prompt_tokens": 24000,
  "reserved_output_tokens": 4000,
  "save": true
}

MCP servers cannot read a client's hidden prompt or chat buffer by themselves. Clients that want Memory Forge to handle active context should call memory_compact before the working context grows too large, then replace the bulky active context with only the returned compacted_context or saved memory reference. Appending compacted context while keeping the original transcript still spends the model window twice.

memory_update

Update content, tags, project, source agent, or importance.

{
  "memory_id": "memory-id",
  "importance": 5,
  "tags": ["api", "database"]
}

memory_forget

Archive by default, or hard-delete only when explicitly requested.

{
  "memory_id": "memory-id",
  "hard_delete": false
}

Development

uv run pytest

Client Integrations

Memory Forge's MCP server can only receive context a client explicitly sends. A richer client integration can handle more memory sources, such as:

• configuring the client to disable duplicate built-in memory;
• installing Memory Forge as the MCP memory backend;
• importing accessible chat history when the user explicitly opts in;
• indexing selected local files or repositories into compact memories.

These integrations should be explicit, local-first, and reversible. Memory Forge should not silently ingest private chats, secrets, or entire filesystems.

Codex

memory-forge-setup codex updates the local Codex config to:

• disable built-in Codex memory;
• register Memory Forge as the memory_forge MCP server;
• point MEMORY_FORGE_DB at the local Memory Forge database.

Use --from-checkout PATH while developing from a local checkout. Use --dry-run to inspect the diff and --check in automation.

Next Clients

Claude Code is the next planned setup target. Its installer should follow the same pattern: configure Memory Forge as the memory backend, avoid duplicate built-in memory where possible, and keep imports or local indexing opt-in.

Roadmap

• Export and import memories as JSONL.
• Optional semantic search using local embeddings.
• Memory compaction and conflict detection.
• Claude Code setup helper.
• Opt-in Codex history import and selected local-file indexing.