Skip to main content
Glama
MikSkrzyp

identity-storage-mcp

by MikSkrzyp
OverviewSchemaSourceRelated ServersDiscussionsAdmin
Python
Local

identity-storage

Portable, auditable long-term memory for AI agents. Runs as a local MCP server backed by a single SQLite file. Agents recall memories through MCP tools; a Stop hook stores session transcripts automatically — no agent discipline required.

Why

Agents like Claude Code are stateless between sessions. identity-storage gives them a memory that survives restarts and stays fully inspectable — no ORM, no migration framework, no hidden state. Point sqlite3 at the file and read everything.

Related MCP server: recollect

Install

The package is not on PyPI yet. Install directly from GitHub:

pip install git+https://github.com/MikSkrzyp/identity-storage-mcp.git

Or run it without installing:

uvx --from git+https://github.com/MikSkrzyp/identity-storage-mcp.git identity-storage-mcp

This installs one console script:

  • identity-storage-mcp — the MCP server (agent calls tools through it)

Configure Claude Code

1. Add the MCP server

claude mcp add identity-storage -s user -- uvx --from git+https://github.com/MikSkrzyp/identity-storage-mcp identity-storage-mcp

2. Add memory instructions to CLAUDE.md

Add this to ~/.claude/CLAUDE.md (global, all projects) or your project's CLAUDE.md:

# Memory — MANDATORY

identity-storage MCP is connected. Follow these rules EVERY session:

1. SEARCH: Call memory_search when the user references past work or you need
   context from previous sessions. Pass the user's prompt as query.

2. STORE: Call memory_store after every non-trivial turn:
   - episodic: events/actions (fixed bug, refactored module, user asked for X)
   - semantic: durable facts (user preferences, project info, tech stack)
   - procedural: how-tos (commands, steps, procedures)
   One memory per distinct thing. Skip idle chat.

3. SESSION END: When the user says exit/quit, store anything not yet saved.

Forgetting to store = permanent loss of the session.
Forgetting to search = working blind.

Configure opencode

1. Add the MCP server

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

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "identity-storage": {
      "type": "local",
      "command": [
        "uvx",
        "--from",
        "git+https://github.com/MikSkrzyp/identity-storage-mcp",
        "identity-storage-mcp"
      ]
    }
  }
}

2. Add memory instructions to AGENTS.md

Add the same memory instructions (from the Claude Code section above) to ~/.config/opencode/AGENTS.md (global) or your project's AGENTS.md.

Tools

The agent sees three tools, each scoped by memory_type (episodic, semantic, procedural, personality, emotional):

Tool

Purpose

memory_search

Full-text search via FTS5 — when the user references past work

memory_store

Store a memory with type classification (episodic/semantic/procedural)

memory_recall

Browse by type, tags, and time window (newest first)

See docs/usage.md for the full input/output schemas.

Configuration

Env var

Default

Purpose

IDENTITY_STORAGE_DB

~/.identity-storage/memory.db

SQLite database file path

The parent directory is created on first run. The schema is applied idempotently on every start, so pointing at a fresh path is safe.

Audit

The database is a regular SQLite file. Read it while the server runs (WAL mode allows concurrent reads):

sqlite3 ~/.identity-storage/memory.db
SELECT id, created_at, content FROM memory
WHERE type='episodic'
ORDER BY created_at DESC;

SELECT * FROM memory
WHERE EXISTS (SELECT 1 FROM json_each(tags) WHERE value='auth');

SELECT m.*
FROM memory m
JOIN memory_fts f ON f.rowid = m.rowid
WHERE f.content MATCH 'auth bug'
ORDER BY rank;

The schema lives in schemas/schema.sql and is the single source of truth. Run .schema in the sqlite3 CLI to see exactly what is in the file.

Other clients

Claude Code and opencode are supported. Both use the same MCP server and the same memory database. For other MCP-compatible clients (Codex, Cursor, etc.), add the MCP server per their docs and add the memory instructions to their equivalent of CLAUDE.md (e.g. .cursorrules for Cursor).

Documentation

Status

Alpha. The MCP contract and the SQLite schema are stable for the episodic case. Semantic memory, procedural memory, consolidation, and embeddings are planned — see docs/architecture.md for the roadmap shape.

License

MIT

Available Tools

3 tools
memory_recallA

Browse memories of one type, newest first. Filter by tags and time window. Use for 'what did I do recently' or 'what happened in this session'. Not for per-turn recall — use memory_search for that.

ParametersJSON Schema
NameRequiredDescriptionDefault
inputYes

Output Schema

ParametersJSON Schema
NameRequiredDescription
recordsYes

TDQS

A4.4/5.0
Behavior4/5

Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?

No annotations are provided, so the description carries full burden. It discloses that the tool browses memories, returns newest first, and supports filtering. It does not explicitly state it is read-only, but that is implied by 'browse'.

Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.

Conciseness5/5

Is the description appropriately sized, front-loaded, and free of redundancy?

Two concise sentences with no superfluous text. The purpose and guidance are front-loaded.

Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.

Completeness4/5

Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?

An output schema exists, so return values are covered. The description provides core usage and differentiation. Missing details on pagination or ordering beyond 'newest first' are minor given schema richness.

Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.

Parameters3/5

Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?

The description mentions 'filter by tags and time window', which adds some context beyond the schema. However, it does not explain the 'limit' or 'memory_type' parameters beyond 'of one type'. With schema description coverage reported as 0%, the description should compensate more.

Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.

Purpose5/5

Does the description clearly state what the tool does and how it differs from similar tools?

The description uses the verb 'browse' with the resource 'memories of one type' and specifies ordering ('newest first'). It clearly differentiates from the sibling 'memory_search' by stating this tool is for browsing and not for per-turn recall.

Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.

Usage Guidelines5/5

Does the description explain when to use this tool, when not to, or what alternatives exist?

Explicit usage guidance is provided: use for 'what did I do recently' or 'what happened in this session', and a direct exclusion: 'Not for per-turn recall — use memory_search for that.'

Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.

memory_storeA

Store a memory. You MUST call this after every non-trivial turn. Choose the type based on what you are saving:

  • episodic: an event that happened — 'fixed the login bug in auth.py', 'user asked for a tic-tac-toe game', 'refactored auth module to use JWT'. Concrete actions and outcomes.

  • semantic: a durable fact that stays true — 'user prefers Python 3.12', 'project uses pytest', 'auth uses JWT', 'user communicates in Polish'. Knowledge about the user or project.

  • procedural: a how-to with steps — 'run tests with pytest -x', 'deploy via npm run build && rsync', 'start dev server: python -m backend.main'. Steps to accomplish something.

Set confidence below 1.0 for inferences, assumptions, or guesses. Use tags for filtering (e.g. project name, topic). Episodic payload keys: session_id, agent, task, outcome, parent_id, metadata. Store one memory per distinct thing. ALWAYS skip idle chat, greetings, and trivial responses. Forgetting to store = permanent loss of the session.

ParametersJSON Schema
NameRequiredDescriptionDefault
inputYes

Output Schema

ParametersJSON Schema
NameRequiredDescription
idYes
created_atYes
memory_typeYes

TDQS

A4.3/5.0
Behavior4/5

Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?

With no annotations, the description reveals important behavior: storing one memory per distinct thing, the consequence of forgetting ('permanent loss'), and specific payload keys for episodic type. It does not mention rate limits or auth, but covers key operational traits.

Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.

Conciseness4/5

Is the description appropriately sized, front-loaded, and free of redundancy?

The description is well-structured with a clear lead sentence and bullet points for types. It is somewhat lengthy but every sentence adds value, avoiding redundancy.

Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.

Completeness4/5

Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?

Given the complexity of the nested input schema and the presence of an output schema, the description is fairly complete. It explains the types, usage rules, and key parameters. It could briefly mention the return value, but the output schema likely covers that.

Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.

Parameters4/5

Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?

Despite the context listing 0% schema description coverage (which seems contradicted by the schema itself), the description adds significant meaning: explaining each memory type, when to set confidence below 1.0, and the role of tags. This goes beyond the schema's brief descriptions.

Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.

Purpose5/5

Does the description clearly state what the tool does and how it differs from similar tools?

The description clearly states 'Store a memory.' and explains when to call it ('after every non-trivial turn'). It distinguishes between memory types (episodic, semantic, procedural) with concrete examples, differentiating from sibling tools like memory_recall and memory_search.

Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.

Usage Guidelines4/5

Does the description explain when to use this tool, when not to, or what alternatives exist?

The description provides explicit guidance on when to use this tool ('after every non-trivial turn'), what types to use, and what to skip ('idle chat, greetings, and trivial responses'). It does not explicitly mention alternatives, but the sibling context is provided separately.

Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

  1. 3 tool updatesv0.1.0
    • First observedmemory_recall
    • First observedmemory_search
    • First observedmemory_store
Deploy Server
One endpoint for every MCP toolRayrun runs every MCP server behind one URL and holds the credentials. Each call is decided per client, tool, and argument, then recorded.ray.runAd

TDQS

A4.4/5.0

Scored across 3 tools

Disambiguation5/5

Each tool has a clearly distinct purpose: memory_recall for browsing by tags/time, memory_search for content search, and memory_store for saving memories. No ambiguity between them.

Naming Consistency5/5

All tools follow a consistent 'memory_' prefix with a verb (recall, search, store), making the tool names predictable and easy to understand.

Tool Count4/5

With 3 tools, the set feels slightly minimal but appropriate for a targeted memory storage and retrieval system. The count is reasonable for the scope.

Completeness3/5

The tool surface covers storing and two retrieval methods, but lacks delete or update operations, which are notable gaps for a complete memory lifecycle.

Maintenance

ActivityStale
ResponsivenessNo issues

Resources

Related MCP Connectors

Related MCP Servers

  • recollect

    danbaileyza
    A
    license
    Not graded
    quality
    C
    maintenance
    Local, cross-agent memory for AI coding agents using a single SQLite file, enabling persistent sessions and durable facts shared across multiple MCP-compatible tools.
    8 npm
    MIT

  • nuzo-memory

    fabionfsc
    A
    license
    Not graded
    quality
    F
    maintenance
    Local-first, auditable memory for AI agents. Provides durable context for MCP hosts with SQLite storage, CLI, and MCP tools for memory management.
    2
    Apache 2.0

  • misMEM

    lucasmella-stack
    A
    license
    Not graded
    quality
    A
    maintenance
    Provides persistent, cooperative memory for LLMs via MCP, with SQLite storage and tools for capturing, recalling, consolidating, crystallizing, and forgetting memories across sessions.
    3
    MIT