Build Persistent Semantic Memory

Mesh Memory

Mesh Memory is a self-hosted semantic memory service with a built-in MCP server. It stores documents (worklogs, decisions, notes, research) in PostgreSQL with pgvector and retrieves them by meaning, so a query like "what database did we pick?" surfaces a saved note that says "chose Redis for caching" even with zero keyword overlap. Embeddings are generated locally with multilingual-e5-base (768 dimensions); the core flow requires no external API keys.

Use this skill when an agent needs persistent memory across sessions: saving its own work, recalling prior decisions, or building a project knowledge base shared between multiple agents.

When to Use This Skill

• Saving a session worklog, decision, or research note so a later session can find it.
• Recalling past work by topic when you do not remember the exact words you used.
• Sharing a long-lived knowledge base across multiple agents, terminals, or teammates.
• Organizing context by role or project through workspaces (one workspace per role/project).
• Looking up structured tags (e.g. all type:decision entries from one project).

Prerequisites

• A running Mesh Memory instance reachable from the MCP server. Local Docker is the common path -- docker compose up -d in the upstream repo brings it up; see https://github.com/dklymentiev/mesh-memory for the full Quick Start.
• The MCP server (mcp_server.py) registered with your client (Claude Code, Cursor, Claude Desktop, or any other MCP-aware agent).
• MESH_API_URL pointing at the running instance (default: http://localhost:8000).

Setup

Register the MCP server in your client configuration:

{
  "mcpServers": {
    "mesh": {
      "command": "python3",
      "args": ["/path/to/mesh-memory/mcp_server.py"],
      "env": {
        "MESH_API_URL": "http://localhost:8000"
      }
    }
  }
}

When the server is reachable, the 13 tools listed below become available.

MCP Tools

Workflows

Save a session worklog

After completing work, persist it for future sessions:

mesh_add(
  content="Investigated 502s on the checkout flow. Root cause: missing CORS header on the cart API. Fix shipped in commit abc123.",
  tags="type:worklog,topic:checkout,date:2026-05-23",
  workspace="developer"
)

date: and source: are added automatically when omitted. Type and topic tags are inferred from nearest neighbors after the embedding completes (5-10 seed documents required before inference kicks in).

Recall past work by meaning

Search across sessions for related context, even with different vocabulary:

mesh_search(query="checkout was failing for some users", limit=5, workspace="developer")

The query shares no keywords with the original note ("502s", "CORS"), but the embedding-based search surfaces it.

Switch role / context

For a multi-role agent, switch the active workspace at the start of a session:

mesh_focus(workspace="sysadmin", prefetch=true, limit=5)

Subsequent calls default to that workspace. Pin a role-prompt document at the top of each workspace so the agent re-orients on every prefetch.

Cross-workspace search with weights

To pull context from related domains without diluting the primary signal:

mesh_search(
  query="nginx rate limit recipe",
  workspaces={"sysadmin": 0.7, "security": 0.2, "developer": 0.1},
  limit=10
)

Results are merged across workspaces and re-scored by workspace weight.

Structured lookups by tag

When you need an exact filter rather than semantic similarity:

mesh_bytag(tags="type:decision,status:active,guid:my-project", limit=20)

Tag Conventions

Mesh accepts arbitrary tags. The recommended prefixes (used by auto-inference and surfaced by mesh_schema):

With fewer than ~5-10 documents in a workspace, neighbor inference is skipped; manually tag seed documents until the corpus self-organizes.

Troubleshooting

Tool calls fail with connection errors. The MCP server cannot reach MESH_API_URL. Verify the instance is up (curl $MESH_API_URL/health returns {"status":"healthy"}) and the env var is set in the MCP config.

A saved document does not appear in semantic search yet. Embedding generation runs in the background. After a save, expect a 1-2 second delay before semantic search hits the new document. mesh_get(guid=...) confirms the document exists immediately.

Search returns results from the wrong domain. The active workspace is not what you expected. Call mesh_focus(workspace="<name>") explicitly, or pass workspace= on every call. With no focus and no explicit param, calls land in the default workspace.

Auto-tagging never adds anything. The workspace has too few documents for neighbor inference (~5-10 minimum). Manually tag a handful of seed documents, then auto-inference takes over.

A deleted document still appears in a search result. Embedding indices are eventually consistent; rerun the search after a few seconds, or use mesh_get(guid=...) to confirm deletion.

Limitations

• Mesh is a knowledge store, not a chat memory. Long conversation transcripts should be summarized before being saved.
• Vector similarity is robust but not perfect; for high-precision structured lookups, prefer mesh_bytag over mesh_search.
• Embeddings run on CPU by default; very large corpora (hundreds of thousands of documents) benefit from a dedicated instance and pgvector tuning, not covered here.
• The optional AI categorizer requires an OpenAI-compatible LLM endpoint and is disabled by default.