MADB — Agent Memory Database

io.github.spshkar84Published by GitHub account spshkar84v0.2.10Updated Oct 24, 2026

Durable causal memory for AI agents — semantic recall, causal lineage, and skills. Local-first.

The publisher's own server: Memory.fit

VerifiedSTDIODesktop onlyKnowledge & MemoryAI & ML

Overview

AI-generated overview

A local, closed-source memory database that gives an AI assistant persistent, causally linked recall of stored memories and reusable skills.

What you can do
Store, look up, search and list memories, with optional cause links, tags and importance scores, and soft-delete them while keeping lineage intact. Recall memories by a composite score of similarity, recency, causal proximity, importance and tag overlap. Walk the causal chain forward or backward from any memory and condense a set of memories into a summary. Save, recall and list reusable skills, and read any memory through a dedicated resource URI. Report engine statistics and workspace capabilities.
When to use it
Useful when you want an assistant to keep durable memory across sessions on your own machine, with links between related events and reusable skills, instead of ad-hoc JSON files or vector stores. Suited to multi-session Claude Code or Claude Desktop setups that should share one memory store.
Requirements
Runs as a local process on the user's machine (desktop only), installed from the PyPI package madb-mcp-server via uvx or pip; Python is needed. Optional environment variables: MADB_DATA_DIR for the data location (default ~/.madb/data), MADB_TENANT_ID to separate memory streams, MADB_TELEMETRY to disable telemetry. No authentication is declared.
Before you install
The engine ships as compiled wheels under a proprietary license and its source is not public. Anonymous count-only usage telemetry is on unless MADB_TELEMETRY is set to off. The server writes memories, skills and deletions to a local store and runs a background daemon on loopback HTTP; the first launch also auto-installs a skill into the user's Claude skills directory.

Installation

In SourceWeft

  1. Open MADB — Agent Memory Database in the dashboard and add it to a workspace.
  2. Enable the server for the chats that should use its tools.

Desktop only via STDIO. STDIO servers start a local process, so they need the SourceWeft desktop host.

Other MCP clients

Follow the launch instructions in the repository.

README

Meta-AgentsDB

Persistent, causal memory for autonomous AI agents. Purpose-built for Claude and the autonomous-agent era.

Meta-AgentsDB (MADB) is an agent-native database — a hybrid LSM store with a causal DAG, composite-score recall, importance-based forgetting, and HLC-ordered MVCC. It is what your agent needs that PostgreSQL + pgvector + Redis + Pinecone cannot give you in one box.

Install (free tier):

bash
claude mcp add madb -- uvx madb-mcp-server      # Claude Code# or:  pip install madb-mcp-server  ·  uvx madb-mcp-server

Memory stays local (~/.madb). Claude Desktop users can install the one-click .mcpb extension. The engine ships as compiled wheels — source is not public. Anonymous usage telemetry can be disabled with MADB_TELEMETRY=off.

What MADB is, in one sentence

The memory substrate Claude reaches for automatically — a focused set of MCP tools, one Skill, one SDK shim — so any Claude session, Claude-Desktop user, or raw-Anthropic-API developer gets persistent, causally-linked, composite-scored agent memory with zero glue code.

How you get MADB

MADB is distributed as compiled binary artifacts under the Meta-Agents.AI Proprietary License. Source code is not public. The three packages:

PackageAudienceHow it's used
meta-agents-db (Python wheel)EveryoneThe Rust engine. Imported by the MCP server and the SDK shim. Direct use optional.
madb-mcp-server (Python wheel)MCP clients — Claude Desktop, Claude Code, Coworkclaude mcp add madb -- uvx madb-mcp-server registers MADB as Claude's memory. Bundles the Skill which auto-installs on first launch.
madb-anthropic (Python wheel)Raw Anthropic-API developers (Python 3.10–3.13)One-import swap: from madb_anthropic import Anthropic. Every messages.create() now auto-persists with causal lineage. No other code changes.

See examples/basic_usage.py for the shim in action.

The MCP tools Claude gets

The free tier exposes a focused core of 13 memory tools:

ToolPurpose
rememberStore a memory with optional caused_by, tags, and importance
recallComposite-scored semantic recall (similarity + recency + causal proximity + importance + tag overlap)
get_memoryPoint lookup by event_id
forgetTombstone soft-delete (preserves DAG lineage)
searchStructured query by tenant + scope + tags
list_recentLast N memories, newest-first
save_skillStore a reusable pattern (idempotent by name, GC-protected)
recall_skillSemantic skill recall
list_skillsAll skills ranked by composite score
trace_causeWalk the causal DAG forward/backward from any memory
summarizeCondense a set of memories into a compact summary
statsEngine metrics snapshot
capabilitiesWhat MADB unlocks in this workspace + entitlement summary

Plus the MCP resource madb://memories/{event_id} for reading any memory as a first-class URI.

The Skill is the reflex

The MCP tools are necessary but not sufficient for "default." Without a Skill telling Claude to prefer MADB, Claude still invents per-task memory schemes — JSON files, SQLite tables, pickle dumps, flat vector stores, localStorage, conversation-history arrays.

skills/madb-memory/SKILL.md routes those anti-patterns to MADB by name. The Skill ships bundled with the MCP server wheel and auto-installs to ~/.claude/skills/madb-memory/ on first launch. It is published under CC-BY-ND 4.0 — share verbatim, don't modify.

A companion skills/madb-onboarding/SKILL.md handles the first run: it detects an empty store and offers a 60-second guided tour (see onboarding/) so your first session ends with a real recall + trace_cause moment instead of an empty store.

One brain, many sessions (0.2.10)

The MADB store is single-process-exclusive by design — one writer, no lock contention, no corruption. But Claude Code spawns one MCP server per session, and you run more than one session. Since 0.2.10 that just works: the wheel bundles a local daemon that becomes the store's one owner, and every session's MCP server is a lightweight RPC client of it over token-guarded loopback HTTP.

  • Nothing to install or manage. The first session spawns the daemon on demand (race-safe); later sessions discover and reuse it. When it's been idle for 15 minutes it exits on its own; the next call respawns it.
  • Upgrades hand off cleanly. After you upgrade the wheel, the newer client asks the old daemon to drain and exit, then takes ownership — no stale process squatting the store, no manual unlock.
  • Same memory everywhere. Two terminals, a SessionStart resume hook, and a background agent all read and write the same brain at the same time.

MADB_MODE=embedded restores the pre-0.2.10 single-session direct open; MADB_MODE=daemon makes daemon failures fatal instead of falling back. MADB_DAEMON_IDLE_SEC tunes the idle timeout (≤0 disables).

License

MADB is distributed under the Meta-Agents.AI Proprietary License. Summary:

  • Permitted: install, run, and use in your own infrastructure — including production — for your own users.
  • Not permitted: redistribute, reverse-engineer, offer as a hosted service, or embed in a commercial product you sell to third parties.
  • Available by separate written agreement: source access, OEM/embed rights, hosted-service rights.

Prior versions of portions of this software were made available under Apache License 2.0. Those revisions remain available under those terms for the specific versions actually so distributed. All new releases are under the Proprietary License only.

The Skill is separately licensed under CC-BY-ND 4.0.

Positioning

MADB is the memory substrate for Claude's path to autonomy. The engine is closed-source and commercial by design; the Skill is open and free to use; the interface is standard MCP. This shape lets Claude adopt MADB as a default capability without Anthropic having to host infrastructure or take on upstream dependencies.

For partnership discussion, source-access review under NDA, or OEM licensing, see Contact below.

Contact

Meta-Agents.AI © 2026 Pushkar Singh. All rights reserved.

Source: README.md at commit 9ff2217

Tools

0
SourceWeft has not started this package in its sandbox yet, so its tools are not known.

Version history

1
  1. v0.2.10LatestSep 16, 2026

Score

40% of the applicable checks met

How complete this listing is, computed from the registry entry, the endpoint and public sources. Not a rating of the server and not a statement that it is safe or working today: reachability is a dated check, and no installs or reviews go into the score.

  • Reached in the last 30 days25 pts

    Keep the endpoint answering the MCP handshake; SourceWeft probes it every 30 days.

  • Auth declared10 pts

    Declare the credential headers in the registry entry, or keep the endpoint answering the probe so its auth can be classified.

  • Tools listed15 pts

    Answer tools/list without a credential, or ship a package that starts in a sandbox without one.

  • DescriptionRequired15 pts
  • Logo10 pts
  • Documentation15 pts
  • License10 pts

    Declare a license in the repository or the registry entry.

Other Memory integrations

Memory.fit

fit.memoryOfficial publisher

Hosted Memory.fit server that lets an AI read your training, sleep, recovery and nutrition data and write back weekly plans, targets and meal plans.

Streamable HTTPWeb executableAuth
Location & LifestyleCredential requiredOct 29, 2026

pg.ddx.io: the PostgreSQL community's memory

io.ddx.pgOfficial publisher

pg.ddx.io: the PostgreSQL community's memory's hosted MCP server, published by pg.ddx.io; no sign-in required; Streamable HTTP; 29 tools.

Streamable HTTPWeb executable
Databases29 toolsOct 24, 2026

var.gg Naming Memory

gg.varOfficial publisher

Remote MCP server that suggests and searches code identifier names from domain concepts in Korean, Japanese, Chinese, Vietnamese and English, and exposes ERD…

Streamable HTTPWeb executable
Databases7 toolsOct 27, 2026

Counterparts

ai.counterparts

Local-first long-term memory for Claude Code and Claude Desktop: the assistant writes notes about you and your projects and recalls them in later sessions.

STDIODesktop only
AI & MLNot validated yetNov 7, 2026

LAHZE - Persian AI Memory Platform

Published by GitHub account pedramgholizadeh

LAHZE - Persian AI Memory Platform's hosted MCP server; no sign-in required; Streamable HTTP; 13 tools.

Streamable HTTPWeb executable
Location & Lifestyle13 toolsNov 3, 2026