server-memory

Local-first durable memory for MCP agents, backed by SQLite and FTS5.

Why · How it works · Installation · Quick start · Tools · Evaluation · Validation · Security

server-memory is an open-source Model Context Protocol server for durable AI-agent memory. It stores project facts, decisions, observations, relations, preferences, and activity in local SQLite databases, then returns compact, scoped context when an agent needs continuity across sessions.

It is designed for agents that repeatedly work on the same repositories, systems, incidents, or long-running tasks and need to remember what was already learned without replaying an entire conversation or loading the full knowledge graph every turn.

It is intentionally boring where memory should be boring: local storage, explicit tools, inspectable data, bounded output, and predictable failure modes.

Data remains local unless it is explicitly exported. The default transport is stdio. An optional shared HTTP daemon binds to localhost and uses local bearer-token authentication by default.

LLM agents commonly lose useful state between sessions. Common workarounds have real costs:

• repeating repository discovery and diagnostics
• pasting large handoff summaries into every new session
• consuming context with stale or irrelevant history
• forgetting accepted decisions, constraints, and unresolved work
• mixing user preferences with project-specific facts
• depending on hosted memory services for data that should stay local

server-memory addresses those problems with durable, queryable memory that can be read selectively instead of replayed wholesale.

The project is intended to improve:

• Cross-session continuity: retain facts and decisions after the original conversation ends.
• Context efficiency: return compact, relevant snippets instead of the entire stored graph.
• Task completion: help agents continue prior work without rediscovering established state.
• Reduced repeated work: preserve attempted commands, known failures, file locations, and next steps.
• Safer scope separation: keep workspace memory distinct from optional global preference memory.
• Local control: use inspectable SQLite databases without requiring external service credentials.

These are design goals, not performance claims. Verified results will be published only after controlled benchmark runs are complete.

The server stores:

• Entities: projects, files, modules, services, people, configurations, incidents, and other named objects.
• Observations: durable facts, decisions, preferences, paths, dependencies, code snippets, and configuration details.
• Relations: typed links between entities.
• Tags: project scopes, pinned items, preferences, and workflow labels.
• Activity: decisions, changes, fixes, and other events worth carrying into later sessions.

Routine recall uses memory_context:

1. Scope the lookup to the active workspace and optional global preference database.
2. Collect candidates through FTS5, optional embeddings, activity links, and fallback matching.
3. Rank candidates using exact-name, lexical, semantic, importance, confidence, pinned, activity, access-recency, and staleness signals.
4. Suppress duplicate or low-value matches.
5. Return bounded snippets plus conflict and stale-state indicators.

The goal is to return the smallest useful memory slice for the current task, not to place the entire database in model context.

• Local-first: core operation requires no hosted database or external service credential.
• Selective recall: query relevant memory rather than replaying all stored history.
• Bounded context: token budgets and compact formatting limit retrieval output.
• Explicit durability: agents choose what to store through MCP tools.
• Inspectable state: memory remains readable, exportable, and testable.
• Scope safety: destructive operations reject scope="all".
• Graceful degradation: lexical retrieval remains available without embeddings.

┌────────────┐         stdio         ┌─────────────────────┐
│ MCP client │ ────────────────────> │    server-memory    │
└────────────┘                       │    FastMCP server   │
                                     ├─────────────────────┤
                                     │ Workspace SQLite DB │
                                     │                     │
                                     │ Global preferences  │
                                     │ DB, when enabled    │
                                     └─────────────────────┘

┌────────────┐      stdio       ┌─────────────────────┐
│ MCP client │ ───────────────> │ server-memory-proxy │
└────────────┘                  └──────────┬──────────┘
                                         │
                                         │ HTTP
                                         │ 127.0.0.1:8765/mcp
                                         ▼
                              ┌─────────────────────┐
                              │ server-memory-serve │
                              │ FastMCP daemon      │
                              └─────────────────────┘

Use shared mode when multiple local clients should access one database process rather than opening the same database independently.

• Python 3.10 or newer
• SQLite with FTS5 enabled
• Git and pip for installation from the repository

Check FTS5 support:

python -c "import sqlite3; c=sqlite3.connect(':memory:'); c.execute('CREATE VIRTUAL TABLE t USING fts5(content)'); c.close(); print('FTS5 available')"

No hosted CI status is used as proof of compatibility. Validate the package locally on the operating system and Python version where it will run.

python -m pip install "server-memory @ git+https://github.com/MK-986123/server-memory.git"

python -m pip install "server-memory[embeddings] @ git+https://github.com/MK-986123/server-memory.git"

Embeddings are optional. Core storage and FTS5 retrieval work without them.

git clone https://github.com/MK-986123/server-memory.git
cd server-memory

python -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -e ".[dev]"

On Windows PowerShell:

.venv\Scripts\Activate.ps1

Install development and embedding dependencies together:

python -m pip install -e ".[dev,embeddings]"

AI coding agents working in this repository should follow AGENTS.md. Contributors should also read CONTRIBUTING.md.

server-memory

Equivalent module form:

python -m server_memory

MEMORY_DB_PATH=<PROJECT_ROOT>/memory.db server-memory

Use the equivalent environment-variable syntax for your shell on Windows.

server-memory-serve \
  --host 127.0.0.1 \
  --port 8765 \
  --transport streamable-http

server-memory-proxy --url http://127.0.0.1:8765/mcp

{
  "mcpServers": {
    "server-memory": {
      "command": "server-memory",
      "env": {
        "MEMORY_PROJECT": "<PROJECT_NAME>"
      }
    }
  }
}

Start server-memory-serve separately, then configure the MCP client to launch the proxy:

{
  "mcpServers": {
    "server-memory": {
      "command": "server-memory-proxy",
      "args": [
        "--url",
        "http://127.0.0.1:8765/mcp"
      ]
    }
  }
}

Use memory only when prior state may materially improve the task.

• Call memory_context(hint="current topic", limit=3-5) when earlier decisions, project facts, preferences, or unresolved work may matter.
• Skip memory lookup for one-off answers or tasks already fully grounded in the current context.
• Store durable facts and decisions, not routine conversation.
• Use log_activity after meaningful changes, fixes, or decisions.
• Tag only facts that must remain prominent as pinned.
• Use explicit workspace or global scope for destructive operations.

server-memory registers MCP tools only. It does not register resources or prompts.

Preference-tagged writes can automatically route to the global database when global preference routing is enabled.

Write tools modify the selected SQLite database. backup_memory writes a database backup. export_graph may expose sensitive memory content, so review exports before sharing them.

Configuration is environment-driven. Empty path overrides in .env.example use platform defaults.

The repository includes deterministic retrieval scenarios for memory_context, including exact-name lookup, importance ranking, pinned facts, access recency, activity links, file-path hints, stale-fact demotion, lexical fallback, and duplicate suppression.

Those tests validate expected ranking behavior, but they do not establish real-world improvements in agent completion rate or token use.

A separate controlled protocol is provided in docs/BENCHMARK_PROTOCOL.md. It compares:

1. fresh sessions with no memory
2. fresh sessions with a token-matched manual handoff summary
3. fresh sessions using server-memory

The protocol measures:

• task completion rate
• durable-fact recall and contradiction rate
• total tokens and tokens to first correct action
• repeated work
• tool-call efficiency
• hit@1, hit@3, and reciprocal rank
• memory latency and end-to-end duration
• stale-memory, leakage, duplicate, and incorrect-write failures

Hosted GitHub Actions are not currently treated as an active validation source for this repository. Workflow definitions may remain under .github/workflows for future use, but this README does not claim that those jobs are running or passing.

Install development dependencies:

python -m pip install -e ".[dev]"

Run the required local checks:

python -m compileall -q src tests scripts
python -m ruff check .
python -m pytest -q

Run the full package and supply-chain checks before a release or substantial pull request:

rm -rf dist build
python -m build
python -m twine check dist/*
python scripts/inspect_wheel.py dist
python -m pip_audit
python scripts/smoke_stdio.py server-memory
server-memory-serve --help
server-memory-proxy --help

On PowerShell, remove build artifacts with:

Remove-Item -Recurse -Force dist, build -ErrorAction SilentlyContinue

The stdio smoke test sends an MCP initialize request to the installed entry point and fails if stdout contains non-protocol output.

When reporting validation, include the exact commands, Python version, operating system, and full failure output. Do not describe a check as passing unless it was actually executed.

See CONTRIBUTING.md for contribution guidance and AGENTS.md for repository-specific agent instructions.

• The stdio server writes protocol data to stdout. Diagnostics should go to stderr or logs.
• The shared HTTP daemon defaults to 127.0.0.1 and local bearer-token authentication.
• The bearer token is generated locally and stored under a platform-native runtime directory unless MEMORY_AUTH_TOKEN_PATH is set.
• No external service credentials are required for the core server.
• Optional embeddings may load local or cached model files depending on the environment and installed extras.
• Review exported graph content before sharing it.
• Do not commit live memory databases, token files, or backups.

Report vulnerabilities through GitHub private vulnerability reporting when available. Do not include secrets or private memory exports in public issues.

See SECURITY.md for the project security policy.

Licensed under the MIT License.