Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    opencode-memory: Local Semantic Memory for AI Agents

    1

    Give your coding agent a persistent SQLite-backed memory for architecture decisions, preferences, and project history.

    $9

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned
    opencode-memory: Local Semantic Memory for AI Agents

    opencode-memory: Local Semantic Memory for AI Agents

    Example session with this skill installed

    Remember that we decided to use repository-local memory for architecture decisions so they stay with the project if the folder moves. Tag it with architecture and memory.

    • Read your context and instructions
    • Compiled the opencode-memory

    Saved the decision in project memory and verified that hybrid search returns it.

    Tool response: Memory stored successfully (hash: 60740f06f791b30ff286706c867cfb71268cba24e04ed7e2810d3428e596f5ff)

    FieldValue
    ScopeProject (memory-project)
    Storage<repository>/.opencode-memory/project.db
    Typedecision
    Tagsarchitecture, memory
    Hash60740f06f791b30ff286706c867cfb71268cba24e04ed7e2810d3428e596f5ff

    Stored content

    Decision: use repository-local memory for architecture decisions so they remain attached to the project when its directory moves.

    Recall verification

    I then searched for “architecture decisions when a project directory moves” using hybrid mode with both tags. The search returned one memory with the same hash and content, confirming that the decision is searchable.

    Next steps

    • In a later session, ask “What did we decide about repository-local architecture memory?” to recall it.
    • If the decision changes, I can replace it with memory_update using versioned: true, preserving the supersession history.
    • Use memory_health to check database health and memory_graph to explore related decisions.

    Connects securely to your tools. The creator never sees your data.

    About this skill

    Stop re-explaining project decisions every time a new coding session starts. opencode-memory gives agents a durable, searchable memory of decisions, preferences, conventions, and confirmed fixes.

    This companion skill teaches agents how to install, configure, and use the open-source opencode-memory MCP server. Your one-time purchase includes the setup guide, exact tool workflows, and memory-management playbook. The server is a separate free MIT-licensed download.

    What the skill teaches agents to do

    • Save atomic facts, architectural decisions, user preferences, project conventions, and incident resolutions.
    • Keep repository facts in project memory and cross-project preferences in global memory.
    • Search by meaning, literal text, tags, and dates. Hybrid mode combines vector retrieval with FTS5 BM25 keyword ranking.
    • Version changed facts and hide superseded memories from default searches.
    • Explore neighbors, shortest paths, subgraphs, shared-neighbor candidates, and unlinked semantic suggestions.
    • Preview filtered deletions before making changes and keep secrets and transient noise out of storage.
    • Report failures honestly instead of claiming a memory was saved or found when the server was unavailable.

    The server is one Rust binary using SQLite, sqlite-vec, and local llama.cpp embeddings.

    What your purchase includes

    • A SKILL.md operating playbook for recall, storage, correction, deletion, and graph exploration.
    • A setup guide with concrete configuration for opencode, Claude Code, and Codex CLI.
    • Generic MCP stdio instructions for other compatible clients.
    • An exact reference for all eight MCP tools, including arguments, filters, versioning, and graph actions.
    • Privacy and troubleshooting guidance for local and remote embedding configurations.
    • Future updates to this skill published through Agensi.

    You are paying for the companion playbook and its updates. The purchase does not include hosted memory, a managed database, or a support subscription.

    How it works

    1. Install the free opencode-memory binary and llama.cpp.
    2. Connect opencode-memory global, opencode-memory project, or both to an MCP-capable client.
    3. Ask the agent to recall relevant context before a task or remember a durable fact afterward.
    4. The skill selects the appropriate scope, search mode, filters, and tool arguments.
    5. When facts change, the agent versions the existing memory instead of creating contradictory duplicates.

    Why use a memory server instead of a notes file?

    A notes file can hold context, but it does not provide semantic ranking, hybrid retrieval, duplicate detection, scoped stores, versioned corrections, or graph traversal. This skill gives the agent consistent rules for what is worth remembering, what should stay out of memory, and how to retrieve or correct facts without inventing successful operations.

    Use cases

    • Recall why a repository chose a particular architecture or dependency.
    • Carry coding-style and communication preferences between projects.
    • Find a previous incident with a similar database or embedding error.
    • Record a confirmed root cause so another agent does not repeat the investigation.
    • Update an outdated convention while preserving its supersession history.
    • Explore memories associated with a design decision before changing it.

    Frameworks and tools

    Rust, Model Context Protocol, SQLite, sqlite-vec, FTS5, llama.cpp, and Qwen3-Embedding-4B by default. The server exposes eight MCP tools for storage, search, listing, updates, deletion, health, telemetry, and graph queries.

    Requirements and known limitations

    • The MCP server is not embedded in the skill ZIP. Install the free open-source binary separately.
    • macOS and Linux are supported. Windows requires WSL; there is no native Windows build.
    • llama.cpp's llama-server must be available on PATH.
    • The default 2560-dimensional Qwen3-Embedding-4B model is downloaded on first use and requires disk space and network access.
    • Memory content stays local with the default embedding configuration. A remote embedding provider, if configured, receives submitted memory and query text.
    • Concrete client configuration is supplied for opencode, Claude Code, and Codex CLI. Other MCP clients can use the generic command, but client-specific setup is not verified by this package.
    • Association-graph traversal reflects periodic graph consolidation rather than real-time edge creation.
    • This is local memory, not a hosted service or a synchronization system between machines.

    Requires opencode-memory MCP server installed on macOS, Linux, or WSL. No support for native Windows outside of WSL. Cannot store sensitive secrets or massive files.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before any payment.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Money-back guarantee

    Doesn't do what its page promises? You get your money back.

    Trust & safety

    Security scan

    Security scan passed 2 days ago

    Requires

    llama.cpp with llama-server on PATH
    macOS or Linux, or Windows through WSL
    Opencode Memory MCP

    Permissions

    Terminal / Shell
    Read Files
    Write Files
    Network Access

    File Scopes

    ~/.config/opencode/memory/**
    **/.opencode-memory/**
    ~/.config/opencode/opencode.jsonc
    ~/.codex/config.toml
    ~/.cache/huggingface/**

    Memory content and embeddings stay local by default. Network access is used to install the server and download the embedding model. Remote embedding endpoints are optional. Installation and configuration changes require user approval. No external accounts are required for the default local setup.

    Listed2 days ago

    What's inside

    Frequently Asked Questions

    Chat history is session-scoped. opencode-memory stores selected facts in SQLite. Global memories cover cross-project preferences; project memories stay attached to their repository and remain available in later sessions.

    The instructions are agent-neutral. The package includes concrete setup for opencode, Claude Code and Codex CLI. Other agents need local MCP stdio support; their client-specific setup has not been verified. Use macOS, Linux, or Windows through WSL.

    Yes. Use global memory for cross-project preferences and project memory for repository decisions. The server keeps them in separate databases, and the skill tells the agent which scope to search or update.

    Connect your AI agent to Agensi through MCP, then ask it to use opencode-memory: Local Semantic Memory for AI Agents on your task. Your agent loads the latest approved version each time. Paid skills require a purchase or active subscription. Skill files are not available for download; your licence does not allow you to save, copy or export any part of a skill.

    The skill is for short, standalone text facts and metadata. It tells agents not to store secrets, private keys, access tokens, large logs, whole files, unverified guesses, or sensitive personal data. These are workflow rules, not a built-in secret scanner.

    Install opencode-memory and llama.cpp, connect the global or project MCP commands to your client, and restart it. macOS and Linux are supported; Windows requires WSL. The first semantic operation downloads the default Qwen3 embedding model. Rust is needed only for source or crates.io installs, not prebuilt binaries.

    New in LLM Ops