The server runs automatically when configured in Claude Code. Manual invocation:

# Default (stdio mode for MCP)
memora-server

# With graph visualization server
memora-server --graph-port 8765

# HTTP transport (alternative to stdio)
memora-server --transport streamable-http --host 127.0.0.1 --port 8080

Claude Code

Add to .mcp.json in your project root:

Local DB:

{
  "mcpServers": {
    "memora": {
      "command": "memora-server",
      "args": [],
      "env": {
        "MEMORA_DB_PATH": "~/.local/share/memora/memories.db",
        "MEMORA_ALLOW_ANY_TAG": "1",
        "MEMORA_GRAPH_PORT": "8765"
      }
    }
  }
}

Cloud DB (Cloudflare D1) - Recommended:

{
  "mcpServers": {
    "memora": {
      "command": "memora-server",
      "args": ["--no-graph"],
      "env": {
        "MEMORA_STORAGE_URI": "d1://<account-id>/<database-id>",
        "CLOUDFLARE_API_TOKEN": "<your-api-token>",
        "MEMORA_ALLOW_ANY_TAG": "1"
      }
    }
  }
}

With D1, use --no-graph to disable the local visualization server. Instead, use the hosted graph at your Cloudflare Pages URL (see Cloud Graph).

Cloud DB (S3/R2) - Sync mode:

{
  "mcpServers": {
    "memora": {
      "command": "memora-server",
      "args": [],
      "env": {
        "AWS_PROFILE": "memora",
        "AWS_ENDPOINT_URL": "https://<account-id>.r2.cloudflarestorage.com",
        "MEMORA_STORAGE_URI": "s3://memories/memories.db",
        "MEMORA_CLOUD_ENCRYPT": "true",
        "MEMORA_ALLOW_ANY_TAG": "1",
        "MEMORA_GRAPH_PORT": "8765"
      }
    }
  }
}

Codex CLI

Add to ~/.codex/config.toml:

[mcp_servers.memora]
  command = "memora-server"  # or full path: /path/to/bin/memora-server
  args = ["--no-graph"]
  env = {
    AWS_PROFILE = "memora",
    AWS_ENDPOINT_URL = "https://<account-id>.r2.cloudflarestorage.com",
    MEMORA_STORAGE_URI = "s3://memories/memories.db",
    MEMORA_CLOUD_ENCRYPT = "true",
    MEMORA_ALLOW_ANY_TAG = "1",
  }

All 43 MCP tools register unconditionally, so every agent session is injected with the full ~12,700-token tool schema even when most tools are never called. MEMORA_TOOL_PROFILE exposes a subset per deployment so a gated tool is genuinely absent — missing from tools/list AND undispatchable (call_tool returns unknown-tool, not a hidden execution). The profile is applied and attested at startup; the active profile and exposed tool count are logged to stderr.

• Unset / empty = full. No existing deployment changes behaviour.
• An unknown value aborts startup with a message naming the valid values. It never silently falls back to full — a typo must not re-expose destructive maintenance tools (memory_rebuild_embeddings, memory_delete_batch) to every worker. Fail closed.
• memory_list is in leader but not agent. It was excluded from both while it cost 163-174s on a D1 store against memory_list_compact's 0.22s; #973 fixed that (now ~1.1s). It stays out of agent because a worker's read surface is deliberately narrow, not for speed.
• The leader/agent boundary is data in memora/tool_profile.py (two frozensets). Editing it is one line, not a sweep of 43 decorators.
• The prune deletes from FastMCP's private _tool_manager._tools dict, so memora pins mcp>=1.27,<1.28 (the audited minor) and runs a startup attestation through the low-level registered MCP request handlers (_mcp_server.request_handlers[ListToolsRequest] / [CallToolRequest] — the actual dispatch callable real client requests use, not the FastMCP.list_tools / call_tool Python helpers) that refuses to start if the installed SDK routes listing/dispatch elsewhere (private-implementation drift). The pin is the static guard; the attestation is the runtime backstop. Bumping the upper bound requires re-running tests/test_tool_profile.py.
• Under container deployment the profile is per container while roles are per agent. One container serving a workspace's leader and its workers needs the leader superset; agent would strip create_section/store_document/delete/digest/tags from the leader.
• memora-server (i.e. memora.server.main()) is the sole supported profiled serving path. A direct embedder that imports memora.server.mcp and calls mcp.run() themselves bypasses profiling entirely (the global mcp still holds all 43 tools); embedders who want profiling must call apply_tool_profile themselves or use main().

# Leader deployment — exposes 19 tools
MEMORA_TOOL_PROFILE=leader memora-server

# Agent worker — exposes 12 tools
MEMORA_TOOL_PROFILE=agent memora-server

# Full (default) — all 43 tools, existing behaviour
memora-server

# Typo refuses to start:
# MEMORA_TOOL_PROFILE=agnt memora-server
# Error: unknown MEMORA_TOOL_PROFILE='agnt'; valid values: full, leader, agent

One memora process can serve every workspace. MEMORA_DATABASES is a JSON registry of {name: storage URI}; a client reaches its store at /mcp/<name>. The selector is the URL already in .mcp.json, not a tool argument — an optional db on every tool is 43 chances to forget one, and every miss would write into someone else's store.

Unset MEMORA_DATABASES is the old shape: one backend from MEMORA_STORAGE_URI / MEMORA_DB_PATH, one /mcp. Existing stdio deployments do not change.

Routing (streamable-http only):

The binding is sticky per MCP session, not per request. A session opened on /mcp/alpha and reused against /mcp/beta still resolves to alpha. A client cannot half-switch databases mid-conversation.

Malformed configuration refuses to start (it does not fall through to the legacy database): bad JSON, a non-object, duplicate keys, an empty URI, a name that is not one URL path segment, or MEMORA_DEFAULT_DB missing/unknown when more than one database is listed.

Worked pair — run this, connect to this. A streamable-HTTP listener, not an MCP command entry (that would spawn a stdio child that never speaks MCP on stdio). Credentials live on the server process.

MEMORA_DATABASES='{"memora":"d1://<account-id>/<memora-db-id>","ob1":"d1://<account-id>/<ob1-db-id>"}' \
MEMORA_DEFAULT_DB=memora \
CLOUDFLARE_API_TOKEN='<token>' \
MEMORA_VECTOR_SCAN_PAGE_SIZE=100 \
memora-server --transport streamable-http --host 127.0.0.1 --port 8000 --no-graph

{
  "mcpServers": {
    "memora": {
      "type": "http",
      "url": "http://127.0.0.1:8000/mcp/ob1"
    }
  }
}

Container / proxy variant (this host's usual launcher, not the command above): scripts/memora-instance.sh up myinstance starts the same HTTP server inside a container and puts scripts/memora_proxy.py on 127.0.0.1:<PORT> (8910 for the memora instance). The workspace URL is then http://127.0.0.1:8910/mcp/ob1. See Container Deployment.

A registry may mix d1://, s3://, and local paths; parse_backend_uri dispatches on the scheme.

memory_stats reports the bound database. It returns database (the name this session actually resolved) and database_source (path, registry_default, or unconfigured). A valid-but-wrong name in .mcp.json is otherwise undetectable: every tool works, reads succeed, and writes land silently in another project's store. Call memory_stats and check database against the workspace you meant.

Health of a multi-database process: GET /health is liveness (no database I/O — the only signal a supervisor may restart on). GET /health/db is an alert surface (always HTTP 200; status is ok, degraded, unknown — no snapshot yet, a refresh timed out, or evidence older than max staleness — or error if the registry itself is unusable). GET /health/db/{name} is the workspace-specific probe (200 or 503). Withdrawing the whole process because one store is degraded takes the healthy ones down with it.

With MEMORA_DATABASES unset, a process still binds one database for its lifetime (MEMORA_STORAGE_URI / MEMORA_DB_PATH). That is the original one-store-one-container-one-port shape.

With MEMORA_DATABASES set, one container serves every workspace and clients select a store by URL path (/mcp/<name>). See Multi-database routing. scripts/memora-instance.sh wants one of STORAGE_URI, VOLUME, or MEMORA_DATABASES per instance file (load() requires at least one). If more than one is set, cmd_up uses MEMORA_DATABASES, then STORAGE_URI, then VOLUME.

Dockerfile builds a credential-free image; scripts/memora-instance.sh deploys one instance from instances/myinstance.env (or another named file). The script's runtime CLI is $MEMORA_CONTAINER_BIN (default container — Apple's CLI). Every container operation the script performs honours that override (build, up, status, logs, down). The generated memora_proxy.py process hardcodes container list, which is also why the proxy exists: that runtime reassigns the container's IP on every start.

./scripts/memora-instance.sh build   myinstance   # build the image
./scripts/memora-instance.sh up      myinstance   # run the container
./scripts/memora-instance.sh proxy   myinstance   # render a LaunchAgent + print install commands
./scripts/memora-instance.sh status                # every instance at a glance

Then point the workspace at it — the whole client config, with no secrets in it. A registry instance needs the store in the path (/mcp/<name>); bare /mcp is the registry default:

{"mcpServers": {"memora": {"type": "http", "url": "http://127.0.0.1:8910/mcp/ob1"}}}

Credentials never enter the image, the instance file, or the workspace's HTTP config. They are read at run time from a separate credential config ($CRED_SOURCE — itself a .mcp.json holding only the mcpServers.memora.env block) and injected with -e. If the instance file does not set CRED_SOURCE, the script uses ~/.config/memora/credentials.mcp.json when that file exists, otherwise ~/repos/agentic-box/.mcp.json. Pass through every variable that file defines, not a hand-picked few: a container started with only the embedding keys silently loses memory_absorb's LLM consolidation instead of failing loudly.

Why the proxy exists — read this before deciding you do not need it. The default runtime (Apple's container) reassigns a container's IP on every start, not just on recreate. An MCP client reads its config once at startup, so a moved address does not produce an error: it produces a permanent silent hang. scripts/memora_proxy.py holds a stable 127.0.0.1:<PORT> in front of the moving address and re-resolves per connection.

Two failure modes it distinguishes, which cost an outage to learn:

• The lookup ran and the container is not listed → it really is gone. Refuse.
• The lookup could not run (timeout under host memory pressure) → nothing new is known. Keep serving the last known good address, bounded by MEMORA_PROXY_STALE_GRACE (300s). Conflating the two took every workspace offline while the containers were answering normally on unchanged addresses.

Set MEMORA_TOOL_PROFILE per instance (see Tool Profiles). Note the profile is per container while roles are per agent: if one container serves a workspace's leader and its workers, it needs the leader superset.

Deploy-time script variable (not a memora-server env var — it never reaches the process inside the container):

instances/README.md covers the config fields and launchd/README.md the supervised proxy. REVERT.md documents restoring a workspace to the direct stdio server.

Memora supports three embedding backends:

Embeddings and the LLM are configured separately.

A partial split (only one MEMORA_EMBEDDING_* set) is rejected so one provider’s secret is never sent to another host.

Trap — OpenRouter has no embeddings endpoint. OpenRouter’s catalogue is chat/multimodal only (no embedding models). Do not point the embedding path at OpenRouter via OPENAI_BASE_URL (or a MEMORA base URL). That combination 404s every embed call; without MEMORA_EMBEDDING_STRICT=1 Memora falls back to TF-IDF and keeps answering, so the store fills with keyword bags while looking healthy. OpenRouter remains fine for the LLM only.

Worked example (LLM via OpenRouter, embeddings via Cloudflare Workers AI):

@cf/baai/bge-m3 is 1024-dimensional. Token needs Workers AI permission. Endpoint shape:

https://api.cloudflare.com/client/v4/accounts/<account_id>/ai/v1

{
  "env": {
    "MEMORA_EMBEDDING_MODEL": "openai",
    "OPENAI_API_KEY": "<openrouter-key>",
    "OPENAI_BASE_URL": "https://openrouter.ai/api/v1",
    "MEMORA_LLM_MODEL": "deepseek/deepseek-chat",
    "MEMORA_EMBEDDING_API_KEY": "<cloudflare-api-token-with-workers-ai>",
    "MEMORA_EMBEDDING_BASE_URL": "https://api.cloudflare.com/client/v4/accounts/<account_id>/ai/v1",
    "OPENAI_EMBEDDING_MODEL": "@cf/baai/bge-m3",
    "MEMORA_EMBEDDING_STRICT": "1"
  }
}

What this fix does (no oversell): embeddings and LLM can use different providers; a partial split is rejected; strict mode turns silent degradation into a hard, named failure.

Automatic: Embeddings and cross-references are computed automatically when you memory_create, memory_update, or memory_create_batch.

Manual rebuild required when the store fingerprint changes — not only MEMORA_EMBEDDING_MODEL, but also:

• Embedding endpoint (MEMORA_EMBEDDING_BASE_URL / host)
• Actual model id (OPENAI_EMBEDDING_MODEL, e.g. switching to @cf/baai/bge-m3)
• Vector kind or dimensions (word-key TF-IDF bags vs dense 1024-d; or 384 vs 1024)
• Mixed store (some rows dense, some sparse) — cosine similarity only shares keys, so mixed kinds yield 0.0 recall for old rows

Fingerprint form: backend|model|repr (e.g. openai|@cf/baai/bge-m3|dense:1024). Legacy meta value openai alone is treated as a mismatch.

# After changing embedding model/endpoint, rebuild all embeddings
memory_rebuild_embeddings

# Then rebuild cross-references to update the knowledge graph
memory_rebuild_crossrefs

A built-in HTTP server starts automatically with the MCP server, serving an interactive knowledge graph visualization.

Details Panel

Timeline Panel

Access locally:

http://localhost:8765/graph

Remote access via SSH:

ssh -L 8765:localhost:8765 user@remote
# Then open http://localhost:8765/graph in your browser

Configuration:

{
  "env": {
    "MEMORA_GRAPH_PORT": "8765"
  }
}

To disable: add "--no-graph" to args in your MCP config.

Graph UI Features

• Details Panel - View memory content, metadata, tags, and related memories
• Timeline Panel - Browse memories chronologically, click to highlight in graph
• History Panel - Action log of all operations with grouped consecutive entries and clickable memory references (deleted memories shown as strikethrough)
• Chat Panel - Ask questions about your memories using RAG-powered LLM chat with streaming responses and clickable [Memory #ID] references
• Time Slider - Filter memories by date range, drag to explore history
• Real-time Updates - Graph, timeline, and history update via SSE when memories change
• Filters - Tag/section dropdowns, zoom controls
• Mermaid Rendering - Code blocks render as diagrams

Node Colors

• 🟣 Tags - Purple shades by tag
• 🔴 Issues - Red (open), Orange (in progress), Green (resolved), Gray (won't fix)
• 🔵 TODOs - Blue (open), Orange (in progress), Green (completed), Red (blocked)

Node size reflects connection count.

When using Cloudflare D1 as your database, the graph visualization is hosted on Cloudflare Pages - no local server needed.

Benefits:

• Access from anywhere (no SSH tunneling)
• Real-time updates via WebSocket
• Multi-database support via ?db= parameter
• Secure access with Cloudflare Zero Trust

Setup:

1. Create D1 database:

   Terminal

   Copy

   npx wrangler d1 create memora-graph
   npx wrangler d1 execute memora-graph --file=memora-graph/schema.sql
   

2. Deploy Pages:

   bash

   Copy

   cd memora-graph
   npx wrangler pages deploy ./public --project-name=memora-graph
   

3. Configure bindings in Cloudflare Dashboard:

   • Pages → memora-graph → Settings → Bindings
   • Add D1: DB_MEMORA → your database
   • Add R2: R2_MEMORA → your bucket (for images)

4. Configure MCP with D1 URI:

   config.json

   Copy

   {
     "env": {
       "MEMORA_STORAGE_URI": "d1://<account-id>/<database-id>",
       "CLOUDFLARE_API_TOKEN": "<your-token>"
     }
   }
   

Access: https://memora-graph.pages.dev

Secure with Zero Trust:

1. Cloudflare Dashboard → Zero Trust → Access → Applications
2. Add application for memora-graph.pages.dev
3. Create policy with allowed emails
4. Pages → Settings → Enable Access Policy

See memora-graph/ for detailed setup and multi-database configuration.

Ask questions about your knowledge base directly from the graph UI. The chat panel uses RAG (Retrieval-Augmented Generation) to search relevant memories and stream LLM responses with tool calling support.

• Toggle via the floating chat icon at bottom-right
• Semantic search finds the most relevant memories as context
• Streaming responses with clickable [Memory #ID] references that focus the graph node
• Tool calling — the LLM can create, update, and delete memories directly from chat (e.g., "save this as a memory", "delete memory #42", "update memory #10 with...")
• Works on both the local server and Cloudflare Pages deployment

Configure the chat model:

Requires an OpenAI-compatible API (OPENAI_API_KEY + OPENAI_BASE_URL for local, OPENROUTER_API_KEY secret for Cloudflare). The chat model must support tool use (function calling).

Find and merge duplicate memories using AI-powered semantic comparison:

# Find potential duplicates (uses cross-refs + optional LLM analysis)
memory_find_duplicates(min_similarity=0.7, max_similarity=0.95, limit=10, use_llm=True)

# Merge duplicates (append, prepend, or replace strategies)
memory_merge(source_id=123, target_id=456, merge_strategy="append")

LLM Comparison analyzes memory pairs and returns:

• verdict: "duplicate", "similar", or "different"
• confidence: 0.0-1.0 score
• reasoning: Brief explanation
• suggested_action: "merge", "keep_both", or "review"

Works with any OpenAI-compatible chat API (OpenAI, OpenRouter, Azure, etc.) via OPENAI_BASE_URL. OpenRouter is fine for this LLM path; it does not provide embeddings — configure embeddings separately (see Semantic Search & Embeddings).

Store structured documents (research reports, architecture decisions, post-mortems) as searchable fragment trees:

# Store a markdown document — auto-parsed into typed fragments
memory_store_document(
    content="# Research Report\n\n## Evidence Table\n| Claim | Confidence |\n...",
    document_key="research/memora-enhancements-2026-04-08",
    tags=["memora/research"]
)
# Returns: {root_id: 230, fragment_count: 100, node_map: {claim: [...], plan_item: [...], ...}}

# Retrieve the full document or specific fragment types
memory_get_document(document_key="research/memora-enhancements-2026-04-08")
memory_get_document(document_key="...", node_kinds=["claim"], content_mode="full")

# Delete a document and all its fragments
memory_delete_document(document_key="research/memora-enhancements-2026-04-08")

How it works: The parser splits markdown by structure — tables become individual claims, numbered lists become plan items, URL lists become references, and risk sections become risk fragments. Each fragment is independently searchable via memory_semantic_search while the full document is retrievable as a unit.

Fragment types: claim, plan_item, reference, section_chunk, risk

Integrity guards: Document fragments are protected from accidental modification:

• memory_delete requires force=True for fragments
• memory_merge refuses to merge fragments
• memory_absorb excludes fragments from similarity matching
• memory_find_duplicates and memory_detect_supersessions skip fragments
• Graph UI hides fragments, shows only the document root node

Structured tools for common memory types:

# Create a TODO with status and priority
memory_create_todo(content="Implement feature X", status="open", priority="high", category="backend")

# Create an issue with severity
memory_create_issue(content="Bug in login flow", status="open", severity="major", component="auth")

# Create a section placeholder (hidden from graph)
memory_create_section(content="Architecture", section="docs", subsection="api")

Analyze stored memories and surface actionable insights:

# Full analysis with LLM-powered pattern detection
memory_insights(period="7d", include_llm_analysis=True)

# Quick summary without LLM (faster, no API key needed)
memory_insights(period="1m", include_llm_analysis=False)

Returns:

• Activity summary — memories created in the period, grouped by type and tag
• Open items — open TODOs and issues with stale detection (configurable via MEMORA_STALE_DAYS; memory_insights default 14, graph UI default 30 — same variable, two consumers)
• Consolidation candidates — similar memory pairs that could be merged
• LLM analysis — themes, focus areas, knowledge gaps, and a summary (requires OPENAI_API_KEY)

Manage relationships between memories:

# Create typed edges between memories
memory_link(from_id=1, to_id=2, edge_type="implements", bidirectional=True)

# Edge types: references, implements, supersedes, extends, contradicts, related_to

# Remove links
memory_unlink(from_id=1, to_id=2)

# Boost memory importance for ranking
memory_boost(memory_id=42, boost_amount=0.5)

# Detect clusters of related memories
memory_clusters(min_cluster_size=2, min_score=0.3)

For offline viewing, export memories as a static HTML file:

memory_export_graph(output_path="~/memories_graph.html", min_score=0.25)

This is optional - the Live Graph Server provides the same visualization with real-time updates.

Browse memories directly in Neovim with Telescope. Copy the plugin to your config:

# For kickstart.nvim / lazy.nvim
cp nvim/memora.lua ~/.config/nvim/lua/kickstart/plugins/

Usage: Press <leader>sm to open the memory browser with fuzzy search and preview.

Requires: telescope.nvim, plenary.nvim, and memora installed in your Python environment.