Ogham MCP

Ogham (pronounced "OH-um") -- persistent, searchable shared memory for AI coding agents. Works across clients.

Contents

• Retrieval quality -- 97.2% R@10 on LongMemEval
• The problem
• Quick start
• Installation methods -- Claude Code, OpenCode, Docker, source
• SSE transport -- multi-agent setup
• CLI -- command-line interface
• Configuration -- env vars, embedding providers, temporal search, lifecycle hooks
• MCP tools -- memory, search, graph, profiles, import/export
• Skills -- ogham-research, ogham-recall, ogham-maintain
• Scoring and condensing
• Cross-encoder reranking -- optional FlashRank for self-hosters
• ONNX local embeddings -- BGE-M3 dense + sparse, no API costs
• Database setup -- Supabase, Neon, vanilla Postgres
  • Upgrading from v0.4.x
• Architecture

Retrieval quality

85.8% QA accuracy on the AMB benchmark harness (500 questions, April 2026) -- 429/500 questions answered correctly using GPT-5-mini with reasoning, evaluated by Gemini 2.5 Flash Lite as a strict judge. Retrieval R@10: 99.5%. AMB is the standardised evaluation harness built by the Vectorize team (creators of Hindsight). Thanks to Nicolo and the Vectorize team for making the harness open.

Previously: 91.8% on our internal LongMemEval benchmark pipeline (gpt-5.4-mini reader, rubric judge). The AMB number is lower because AMB uses a stricter substring-matching judge -- see the full write-up for methodology differences.

0.554 nugget score on BEAM 100K (400 questions across 10 memory abilities, ICLR 2026), using the paper's exact judge prompt from Appendix G. The published baseline is 0.358 (Llama-4-Maverick + LIGHT). Retrieval R@10: 0.737. Seven of nine categories beat the paper. Full write-up.

End-to-end QA accuracy on LongMemEval (retrieval + LLM reads and answers):

Retrieval only (R@10 -- no LLM in the search loop):

Other retrieval systems that report similar R@10 numbers typically use cross-encoder reranking, NLI verification, knowledge graph enrichment, and LLM-as-a-judge pipelines. Ogham reaches 97.2% with one Postgres query. Optional FlashRank reranking is available for self-hosters who want extra ranking precision.

These tables measure different things. QA accuracy tests whether the full system (retrieval + LLM) produces the correct answer. R@10 tests whether retrieval alone finds the right memories. Ogham is a retrieval engine -- it finds the memories, your LLM reads them.

Full breakdown: ogham-mcp.dev/features

The problem

AI coding agents forget everything between sessions. Switch from Claude Code to Cursor to Kiro to OpenCode and context is lost. Decisions, gotchas, architectural patterns -- gone. You end up repeating yourself, re-explaining your codebase, re-debugging the same issues.

Ogham gives your agents a shared memory that persists across sessions and clients.

Quick start

1. Install

uvx --from ogham-mcp ogham init

This runs the setup wizard. It walks you through everything: database connection, embedding provider, schema migration, and writes MCP client configs for Claude Code, Cursor, VS Code, and others.

You need a database before running this. Either create a free Supabase project or a Neon database. The wizard handles the rest.

Using Neon or self-hosted Postgres? Install with the postgres extra so the driver is available:

uvx --from 'ogham-mcp[postgres]' ogham init

2. Add to your MCP client

The wizard configures everything and writes your client config -- including all environment variables the server needs. For Claude Code, it runs claude mcp add automatically. For other clients, copy the config snippet it prints.

3. Use it

Tell your agent to remember something, then ask about it later -- from the same client or a different one. It works because they all share the same database.

Manual setup (if you prefer)

If you'd rather configure things yourself instead of using the wizard:

# Supabase
export SUPABASE_URL=https://your-project.supabase.co
export SUPABASE_KEY=your-service-role-key
export EMBEDDING_PROVIDER=openai  # or ollama, mistral, voyage
export OPENAI_API_KEY=sk-...      # for your chosen provider

# Or Postgres (Neon, self-hosted)
export DATABASE_BACKEND=postgres
export DATABASE_URL=postgresql://user:pass@host/db
export EMBEDDING_PROVIDER=openai
export OPENAI_API_KEY=sk-...

Run the schema migration (sql/schema.sql for Supabase, sql/schema_postgres.sql for Neon/self-hosted), then add the MCP server to your client.

Installation methods

Claude Code

claude mcp add ogham -- uvx ogham-mcp

OpenCode

Add to ~/.config/opencode/opencode.json:

{
  "mcp": {
    "ogham": {
      "type": "local",
      "command": ["uvx", "ogham-mcp"],
      "environment": {
        "SUPABASE_URL": "https://your-project.supabase.co",
        "SUPABASE_KEY": "{env:SUPABASE_KEY}",
        "EMBEDDING_PROVIDER": "openai",
        "OPENAI_API_KEY": "{env:OPENAI_API_KEY}"
      }
    }
  }
}

Docker

docker run --rm \
  -e SUPABASE_URL=https://your-project.supabase.co \
  -e SUPABASE_KEY=your-key \
  -e EMBEDDING_PROVIDER=openai \
  -e OPENAI_API_KEY=sk-... \
  ghcr.io/ogham-mcp/ogham-mcp

From source

git clone https://github.com/ogham-mcp/ogham-mcp.git
cd ogham-mcp
uv sync
uv run ogham --help

SSE transport (multi-agent)

By default, Ogham runs in stdio mode -- each MCP client spawns its own server process. For multiple agents sharing one server, use SSE mode:

ogham serve --transport sse --port 8742

The server runs as a persistent background process. All clients connect to the same instance -- one database pool, one embedding cache, shared memory.

Client config for SSE (any MCP client):

{
  "mcpServers": {
    "ogham": {
      "url": "http://127.0.0.1:8742/sse"
    }
  }
}

Health check at http://127.0.0.1:8742/health (cached, sub-10ms).

Configure via env vars (OGHAM_TRANSPORT=sse, OGHAM_HOST, OGHAM_PORT) or CLI flags. The init wizard (ogham init) walks through SSE setup if you choose it.

Entry points

Ogham has two entry points:

• ogham -- the CLI. Use this for ogham init, ogham health, ogham search, and other commands you run yourself. Running ogham with no arguments starts the MCP server.
• ogham-serve -- starts the MCP server directly. This is what MCP clients should call. When you run uvx ogham-mcp, it invokes ogham-serve.

CLI

ogham init                      # Interactive setup wizard
ogham health                    # Check database + embedding provider
ogham config                    # Show runtime configuration (secrets masked)
ogham store "some fact"         # Store a memory
ogham search "query"            # Search memories (hybrid: semantic + keyword)
ogham search "q" --json         # JSON output for scripting
ogham search "q" --tags "a,b"   # Filter by comma-separated tags
ogham list                      # List recent memories
ogham list --json               # JSON output
ogham delete <id>               # Delete a memory by ID
ogham use <profile>             # Switch default profile
ogham profiles                  # List profiles and counts
ogham stats                     # Profile statistics
ogham export -o backup.json     # Export memories
ogham import backup.json        # Import memories
ogham cleanup                   # Remove expired memories
ogham hooks install             # Auto-detect client + configure hooks
ogham hooks recall              # Read from the stone (load project context)
ogham hooks inscribe            # Carve into the stone (capture activity)
ogham serve                     # Start MCP server (stdio, default)
ogham serve --transport sse     # Start SSE server on port 8742
ogham openapi                   # Generate OpenAPI spec

Multi-profile search

Search across multiple profiles in a single query (v0.8.5+):

# MCP tool
hybrid_search(query="architecture decisions", profiles=["work", "shared"])

# Python library
from ogham.service import search_memories_enriched
results = search_memories_enriched(
    query="architecture decisions",
    profile="work",
    profiles=["work", "shared", "project-alpha"],
)

When profiles is set, results include memories from all listed profiles with a profile field showing which profile each result came from.

Configuration

Embedding providers

EMBEDDING_DIM must match the vector(N) column in your database schema. The default schema uses vector(512). If you use Mistral, you need to alter the column to vector(1024) before storing anything.

Each provider clusters vectors differently, so the similarity threshold matters. Start with the recommended value and adjust based on your results.

Temporal search

Search queries with time expressions like "last week" or "three months ago" are resolved automatically using parsedatetime -- no configuration needed. This handles roughly 80% of temporal queries at zero cost.

For expressions that parsedatetime cannot parse ("the quarter before last", "around Thanksgiving"), set TEMPORAL_LLM_MODEL to call an LLM as a fallback:

# Self-hosted with Ollama (free, local)
TEMPORAL_LLM_MODEL=ollama/llama3.2

# Cloud API
TEMPORAL_LLM_MODEL=gpt-4o-mini

Any litellm-compatible model string works -- deepseek/deepseek-chat, moonshot/moonshot-v1-8k, etc. The LLM is only called when parsedatetime fails and the query has temporal intent, so costs stay near zero.

If TEMPORAL_LLM_MODEL is empty (the default), parsedatetime handles everything on its own. Requires the litellm package (pip install litellm or install Ogham with the appropriate extra).

Lifecycle hooks

Ogham hooks inject memory context at session start and preserve it across compaction. Install for your client:

ogham hooks install

Two commands, named after the Ogham stones:

• recall -- read from the stone. Searches Ogham for memories relevant to your project and injects them as context. Fires at session start and after compaction.
• inscribe -- carve into the stone. Captures meaningful tool activity as memories. Skips noise (ls, cat, git status) and only stores signal (commits, deploys, errors, config changes). Fires after tool use and before compaction. Secrets are masked before storing.

Smart filtering: Hooks don't capture everything. Routine commands (ls, pwd, git add) are skipped. Only signal events (errors, deployments, commits, config changes) are stored -- typically 20-30 memories per session instead of hundreds.

Secret masking: API keys, tokens, passwords, and JWTs are automatically replaced with ***MASKED*** before storing. The event is captured ("configured Stripe API key") but the actual secret never touches the database.

MCP tools

Memory operations

Search

Knowledge graph

Profiles

Import / export

Maintenance

Skills

Ogham ships with three workflow skills in skills/ that wire up common MCP tool chains. Install them in Claude Code, Cursor, or any client that supports skills.

Skills call existing MCP tools -- they don't replace them. The MCP server must be connected for skills to work.

Install all three with npx:

npx skills add ogham-mcp/ogham-mcp

Or install a specific skill:

npx skills add ogham-mcp/ogham-mcp --skill ogham-recall

Manual install (copy from a local clone):

cp -r skills/ogham-research skills/ogham-recall skills/ogham-maintain ~/.claude/skills/

Scoring and condensing

Ogham goes beyond storing and retrieving. Three server-side features run automatically, no configuration needed.

Novelty detection. When you store a memory, Ogham checks how similar it is to what you already have. Redundant content gets a lower novelty score and ranks quieter in search results. You can still find it, but it won't push out more useful memories.

Content signal scoring. Memories that mention decisions, errors, architecture, or contain code blocks get a higher signal score. A debug session where you fixed a real bug ranks above a casual note about a meeting. The scoring is pure regex, no LLM involved.

Automatic condensing. Old memories that nobody accesses gradually shrink. Full text becomes a summary of key sentences, then a one-line description with tags. The original is always preserved and can be restored if the memory becomes relevant again. Run compress_old_memories manually or on a schedule. High-importance and frequently-accessed memories resist condensing.

Entity enrichment

Every memory is automatically enriched at ingest with structured entity tags -- no LLM calls, pure regex and dictionary matching across 18 languages.

Six entity categories. Events (wedding, concert, meeting), activities (hiking, coding, cooking), emotions (frustrated, happy, relieved), relationships (sister, boss, colleague), quantities (3 books, 5 miles), and locations (Berlin, Tokyo -- via GeoNames database).

18 languages. English, German, French, Spanish, Italian, Portuguese, Brazilian Portuguese, Dutch, Polish, Russian, Ukrainian, Turkish, Arabic, Hindi, Japanese, Korean, Chinese, and Irish. Each language includes common inflected forms (case endings, verb tenses, lenition) so "svadʹbu" matches "svadʹba" in Russian and "bhainis" matches "bainis" in Irish.

Timeline table. Search results include a chronological timeline with pre-computed "days ago" and memory ID cross-references. Helps LLM readers answer temporal questions without doing date arithmetic.

Lost in the Middle reordering. Search results are reordered so the highest-relevance memories appear at the start and end of the context, where LLMs pay the most attention (Liu et al., 2023).

Cross-encoder reranking

Optional FlashRank cross-encoder reranking for self-hosters who want better ranking precision. Adds ~300ms per search on CPU.

After Ogham's hybrid search returns candidates, FlashRank (ms-marco-MiniLM-L-12-v2, 21MB) rescores each result against the query using deeper token-level attention. The final score blends retrieval ranking with cross-encoder ranking.

BEAM benchmark impact: R@10 0.69 → 0.70, MRR +8pp. Biggest gain: temporal reasoning 0.84 → 0.98. Full results.

Install and enable:

pip install ogham-mcp[rerank]
# or: uv add ogham-mcp[rerank]

export RERANK_ENABLED=true
export RERANK_ALPHA=0.55   # 55% cross-encoder, 45% retrieval score

The model downloads on first use (~21MB). Self-hosters who want speed over precision leave it off (the default).

ONNX local embeddings

Run BGE-M3 locally with ONNX Runtime -- dense and sparse vectors in a single model pass, no API calls, no GPU required. Contributed by @ninthhousestudios.

The ONNX provider produces 1024-dim dense vectors plus neural sparse vectors. When sparse vectors are available, Ogham automatically uses three-signal Reciprocal Rank Fusion (dense + FTS + sparse) instead of the default two-signal path.

Install and configure:

pip install ogham-mcp[onnx]

# Download the model (~2.2GB)
ogham download-model bge-m3

export EMBEDDING_PROVIDER=onnx
export EMBEDDING_DIM=1024

Your database schema must use vector(1024) for the embedding column. Performance on CPU: ~0.3s per short text, ~10s for long documents (5K+ chars). RSS: ~4.3GB peak.

The ONNX provider is designed for self-hosters who want zero API costs. Cloud users should use Gemini or Voyage for lower latency.

Database setup

Ogham works with Supabase or vanilla PostgreSQL. Run the schema file that matches your setup:

Supabase and Neon both include pgvector out of the box -- no extra setup needed. If you're self-hosting Postgres, you need PostgreSQL 15+ with the pgvector extension installed. We develop and test against PostgreSQL 17.

For Postgres, set DATABASE_BACKEND=postgres and DATABASE_URL=postgresql://... in your environment.

Upgrading from v0.4.x

If you already have an Ogham database, run the upgrade script to add temporal columns, halfvec compression, and lz4:

# Postgres / Neon (psql required)
./sql/upgrade.sh $DATABASE_URL

# Or run migrations individually
psql $DATABASE_URL -f sql/migrations/012_temporal_columns.sql
psql $DATABASE_URL -f sql/migrations/013_halfvec_compression.sql
psql $DATABASE_URL -f sql/migrations/014_lz4_toast_compression.sql
psql $DATABASE_URL -f sql/migrations/015_temporal_auto_extract.sql
psql $DATABASE_URL -f sql/migrations/016_sparse_embedding.sql
psql $DATABASE_URL -f sql/migrations/017_rrf_bm25.sql

# Supabase: paste each migration file into the SQL Editor

Migration 016 adds the sparse_embedding column for ONNX BGE-M3 sparse vectors. Migration 017 upgrades the search function to true Reciprocal Rank Fusion with length-normalised keyword scoring -- re-run your schema file to apply.

All migrations are idempotent -- safe to re-run. The upgrade script checks your pgvector version and skips halfvec if pgvector is below 0.7.0.

New installs don't need migrations -- the schema files already include everything.

Architecture

Ogham runs as an MCP server over stdio or SSE. Your AI client connects to it like any other MCP tool.

AI Client (Claude Code, Cursor, Kiro, OpenCode, ...)
    |
    | stdio (MCP protocol)
    |
Ogham MCP Server
    |
    | HTTPS (Supabase REST API) or direct connection (Postgres)
    |
PostgreSQL + pgvector

Memories are stored as rows with vector embeddings. Search combines pgvector cosine similarity with PostgreSQL full-text search using Reciprocal Rank Fusion (RRF) -- position-based, score-agnostic fusion that handles different score scales correctly. Optional FlashRank cross-encoder reranking adds a second pass for self-hosters. The Supabase backend uses postgrest-py directly (not the full Supabase SDK) for a lightweight dependency footprint.

The knowledge graph uses a memory_relationships table with recursive CTEs for traversal -- no separate graph database.

Research foundations

Ogham's retrieval pipeline combines established information retrieval and cognitive science techniques:

• Hybrid search -- Reciprocal Rank Fusion (Cormack, Clarke & Butt, SIGIR 2009) combining dense vector similarity (pgvector) with BM25-style keyword matching (PostgreSQL tsvector). Two independent retrieval systems, rank-fused without score normalisation.
• Entity overlap boost -- memories sharing named entities with the query receive a bounded relevance boost (up to 1.4x), inspired by entity-linking literature (Kolitsas et al., CoNLL 2018). Entity extraction covers 18 languages via YAML-based word lists with no LLM in the write path.
• Matryoshka embeddings -- flexible dimensionality via Matryoshka Representation Learning (Kusupati et al., NeurIPS 2022). Embedding providers (OpenAI, Voyage, Gemini, Ollama) produce native-dimension vectors truncated to 512d, enabling provider-portable storage without re-embedding.
• Temporal diversity re-ranking -- density-gated soft penalty preventing semantic clustering on a single time period, extending Maximal Marginal Relevance principles (Carbonell & Goldstein, SIGIR 1998). Only activates when the top-k results are temporally concentrated, leaving well-distributed results untouched.
• ACT-R importance scoring -- cognitive-architecture-inspired memory weighting based on recency, access frequency, and surprise (Anderson & Lebiere, 1998). Frequently accessed memories stay sharp, rarely accessed ones fade, disputed ones drop in ranking without deletion.
• Read-time fact extraction -- query-aware extraction at retrieval time preserves verbatim storage for auditability, contrasting with write-time compression approaches. Verbatim storage ensures the ground truth is always available for re-extraction with different questions later -- a design choice informed by alignment considerations in persistent agent memory (Anthropic, arXiv:2510.05179).

Documentation

Full docs and integration guides at ogham-mcp.dev.

Credits

Inspired by Nate B Jones and his work on persistent AI memory.

Named after Ogham, the ancient Irish alphabet carved into stone -- the original persistent memory.

License

MIT