allsource-prime v0.22.0 · MCP

Persistent memory for Claude and other MCP clients

Prime stores knowledge-graph relationships, embeddings, provenance, and compressed context in a local event store. Claude Desktop, Claude Code, and other MCP clients read and write that memory directly.

Install Prime MCP setup docs

One memory store across agent surfaces

One cargo install, one config edit per client. The same Prime store serves Claude Desktop, Claude Code, Cursor, and OpenCode — one source of truth, every place your agents work.

1. Install the binary

cargo install allsource-prime

2a. Claude Desktop — ~/Library/Application Support/Claude/claude_desktop_config.json

{
  "mcpServers": {
    "prime": {
      "command": "allsource-prime",
      "args": [
        "--data-dir", "~/.prime/memory",
        "--auto-inject"
      ]
    }
  }
}

2b. Claude Code — one command

claude mcp add prime allsource-prime \
  --data-dir ~/.prime/memory \
  --auto-inject

2c. Cursor — ~/.cursor/mcp.json

{
  "mcpServers": {
    "prime": {
      "command": "allsource-prime",
      "args": [
        "--data-dir", "~/.prime/memory",
        "--auto-inject"
      ]
    }
  }
}

2d. OpenCode — opencode.json

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "prime": {
      "type": "local",
      "command": ["allsource-prime", "--data-dir", "~/.prime/memory", "--auto-inject"],
      "enabled": true
    }
  }
}

Why this works. The --auto-inject flag exposes a prime://auto-context MCP resource. Each client reads it at conversation start — a compressed markdown index of everything you've told it, organized by domain, in ~500–1,000 tokens. No prompt engineering required, and every client sees the same index because they share the same local Prime store (or, if you wire --sync-to, the same hosted one).

Local writer rule. Run one writer process against a local --data-dir at a time. Use separate directories or tenant-scoped hosted persistence for concurrent clients.

19 MCP tools Claude picks automatically

You don't call these. Claude does — the tool descriptions are written for agent consumption. v0.22.0 adds text-only inputs for prime_embed and prime_recall so the model never has to hand over a precomputed vector.

Tool	When Claude calls it
prime_add_node	Create a node — person, concept, project, decision, insight.
prime_add_edge	Connect two nodes with a directed relation (works_on, impacts, depends_on…).
prime_embed	Make a node findable by meaning. Pass text — Prime embeds it server-side via fastembed.
prime_recall	Hybrid recall: vector similarity + graph proximity + temporal recency. Pass text or a vector.
prime_context	Combined retrieval with tiered depth (L0 stats / L1 conversation / L2 full hybrid).
prime_index	Compressed knowledge index — token-efficient markdown of everything you know, by domain.
prime_neighbors	Walk the graph around a node. Multi-hop BFS with direction + relation filters.
prime_search	List all nodes of a given type. Cheap, broad.
prime_shortest_path	Find how two entities are connected — the chain of relationships between them.
prime_similar	Find the most semantically similar embeddings to a stored vector.
prime_history	Full audit trail for any entity — every creation, update, and deletion.
prime_forget	Soft-delete a node. Invisible to queries, preserved in history.
prime_stats	Graph overview: total nodes, edges, types, relations.

Where Prime fits

Prime is a retrieval layer over durable event history, not a general database or a human publishing suite.

Use Prime for agent memory

Choose Prime when agents need graph and vector recall, compressed context, temporal history, and provenance through MCP.

Use Core without Prime for event history

Choose Core alone when replay, point-in-time reconstruction, and projections matter but semantic or graph recall does not.

Use a CMS for human publishing workflows

Choose a CMS when editorial UI, approvals, previews, scheduling, and collaborative page editing are primary requirements.

What's under the hood

Prime is an AllSource event store, not a separate database. Every mutation — node, edge, vector, soft-delete — appends an immutable event to the WAL and is replayed by projections that hold the queryable state. WAL + Parquet durability with CRC32 checksums; Snappy-compressed columnar snapshots; in-memory Core indexed reads measured 11.9µs p99 in the published reference benchmark. Hybrid recall adds graph and vector work and has separate end-to-end latency.

Embeddings are computed in-process via fastembed (AllMiniLML6V2, 384 dims, ~25 MB, auto-downloaded into the fastembed cache on first call). No external embedding service is required. Subsequent embeddings take ~1–3 ms on a modern CPU.

Technical deep-dive crates.io/allsource-prime ↗ Source

Already running Claude?

Then you're halfway there. Install allsource-prime, add the MCP config block, restart your client. Claude starts writing to memory on its own — no prompt changes needed.

Back to install