Skip to main content
Glama

โœจ Features

  • ๐Ÿ” Hybrid search โ€” BM25 (full-text) + vector KNN (pgvector) fused via Reciprocal Rank Fusion

  • ๐Ÿ“‚ Multi-user โ€” each user sees only their own memories; workspaces for shared team knowledge

  • ๐Ÿ› ๏ธ 11 MCP tools โ€” store, retrieve, recall, list, search by tag, get, update, delete, move, list workspaces, dream report

  • ๐Ÿ’ค Dreaming โ€” a guided consolidation pass (dream MCP prompt + dream_report tool): finds near-duplicate and conflicting memories, merges them into insights non-destructively, driven by the client LLM

  • ๐ŸŒ Web UI โ€” browse, search, edit and delete memories in the browser; export/import JSON

  • ๐Ÿ”’ OAuth 2.1 โ€” full in-process authorization server, always enabled

  • ๐Ÿ˜ PostgreSQL โ€” pgvector for embeddings, tsvector for full-text search

  • ๐Ÿ“Š Progressive disclosure โ€” search returns compact snippets by default; call get_memory(name) only for what you need, reducing token usage

  • ๐Ÿ” Deduplication โ€” automatically detects near-identical memories before saving, preventing noise accumulation


Related MCP server: Memory MCP Server

๐Ÿ†š How Memlord compares

Memlord

OpenMemory

mcp-memory-service

basic-memory

Search

BM25 + vector + RRF

Vector only (Qdrant)

BM25 + vector + RRF

BM25 + vector

Embeddings

Local ONNX, zero config

OpenAI default; Ollama optional

Local ONNX, zero config

Local FastEmbed

Storage

PostgreSQL + pgvector

PostgreSQL + Qdrant

SQLite-vec / Cloudflare Vectorize

SQLite + Markdown files

Multi-user

โœ…

โŒ single-user in practice

โš ๏ธ agent-ID scoping, no isolation

โŒ

Workspaces

โœ… shared + personal, invite links

โš ๏ธ "Apps" namespace

โš ๏ธ tags + conversation_id

โœ… per-project flag

Authentication

โœ… OAuth 2.1

โŒ none (self-hosted)

โœ… OAuth 2.0 + PKCE

โŒ

Web UI

โœ… browse, edit, export

โœ… Next.js dashboard

โœ… rich UI, graph viz, quality scores

โŒ local; cloud only

MCP tools

11

5

15+

~20

Self-hosted

โœ… single process

โœ… Docker (3 containers)

โœ…

โœ…

Memory input

Manual (explicit store)

Auto-extracted by LLM

Manual

Manual (Markdown notes)

Memory types

fact / preference / instruction / feedback / decision / insight

auto-extracted facts

โ€”

observations + wiki links

Time-aware search

โœ… natural language dates

โš ๏ธ REST only, not in MCP tools

โ€”

โœ… recent_activity

Token efficiency

โœ… progressive disclosure

โŒ

โ€”

โœ… build_context traversal

Import / Export

โœ… JSON

โœ… ZIP (JSON + JSONL)

โ€”

โœ… Markdown (human-readable)

License

AGPL-3.0 / Commercial

Apache 2.0

Apache 2.0

AGPL-3.0

Where competitors have a real edge:

  • OpenMemory โ€” auto-extracts memories from raw conversation text; no need to decide what to store manually; good import/export

  • mcp-memory-service โ€” richer web UI (graph visualization, quality scoring, 8 tabs); more permissive license (Apache 2.0); multiple transport options (stdio, SSE, HTTP)

  • basic-memory โ€” memories are human-readable Markdown files you can edit, version-control, and read without any server; wiki-style entity links form a local knowledge graph; ~20 MCP tools

When to pick Memlord:

  • You want zero-config local embeddings โ€” ONNX model ships with the server, no Ollama or external API needed

  • You run a multi-user team server with proper OAuth 2.1 auth and invite-based workspaces

  • You want a production-grade database (PostgreSQL) that scales beyond a single machine's SQLite

  • You manage memories explicitly โ€” store exactly what matters, typed and tagged, not everything the LLM decides to extract

  • You want a self-hosted Web UI with full CRUD and JSON export, without a cloud subscription


๐Ÿš€ Quickstart

๐Ÿณ Docker

cp .env.example .env
docker compose up

HTTP server (multi-user, Web UI, OAuth)

# Install dependencies
uv sync --dev

# Download ONNX model (~23 MB)
uv run python scripts/download_model.py

# Run migrations
alembic upgrade head

# Start the server
memlord

Open http://localhost:8000 for the Web UI. The MCP endpoint is at /mcp.


๐Ÿ” How It Works

Each search request runs BM25 and vector KNN in parallel, then merges results via Reciprocal Rank Fusion:

flowchart TD
    Q([query]) --> BM25["BM25\nsearch_vector @@ websearch_to_tsquery"]
    Q --> EMB["ONNX embed\nparaphrase-multilingual-MiniLM-L12-v2 ยท 384d ยท local"]
    EMB --> KNN["KNN\nembedding <=> query_vector\ncosine distance"]
    BM25 --> RRF["RRF fusion\nscore = 1/(k+rank_bm25) + 1/(k+rank_vec)\nk=60"]
    KNN --> RRF
    RRF --> R([top-N results])

โš™๏ธ Configuration

All settings use the MEMLORD_ prefix. See .env.example for the full list.

Variable

Default

Description

MEMLORD_DB_URL

postgresql+asyncpg://postgres:postgres@localhost/memlord

PostgreSQL connection URL

MEMLORD_PORT

8000

Server port

MEMLORD_BASE_URL

http://localhost:8000

Public URL for OAuth (HTTP mode)

MEMLORD_OAUTH_JWT_SECRET

memlord-dev-secret-please-change

JWT signing secret (HTTP mode)

Set MEMLORD_BASE_URL to your public URL and change MEMLORD_OAUTH_JWT_SECRET before deploying.


๐Ÿ› ๏ธ MCP Tools

Tool

Description

store_memory

Save a memory (idempotent by content); raises on near-duplicates; optional expires_at

retrieve_memory

Hybrid semantic + full-text search; returns snippets by default

recall_memory

Search by natural-language time expression; returns snippets by default

list_memories

Paginated list with type/tag filters

search_by_tag

AND/OR tag search

get_memory

Fetch a single memory by name with full content (expired included)

update_memory

Update content, type, tags, metadata, or expiry by name (and optionally rename)

delete_memory

Delete by name

move_memory

Move a memory to a different workspace

list_workspaces

List workspaces you are a member of (including personal)

dream_report

Read-only consolidation candidates: similar memory pairs, expired and expiring-soon memories

The dream MCP prompt walks the client LLM through a full consolidation pass over the dream_report output: classify similar pairs (duplicate / complementary / conflict), merge into insight memories, retire superseded ones via expires_at โ€” never destructively.

Workspace management (create, invite, join, leave) is handled via the Web UI.


๐Ÿ’ป System Requirements

  • Python 3.12

  • PostgreSQL โ‰ฅ 15 with pgvector extension

  • uv โ€” Python package manager


๐Ÿ‘จโ€๐Ÿ’ป Development

pyright src/           # type check
ruff format .          # format
pytest                 # run tests
alembic-autogen-check  # verify migrations are up to date

๐Ÿ“„ License

Memlord is dual-licensed:

Available Tools

11 tools
delete_memoryA
Destructive

Delete a memory by name. Pass workspace to disambiguate if the name exists in multiple workspaces.

Inputs: name (required), workspace. Returns: name, success

dream_reportA
Read-only

Candidates for memory consolidation: similar pairs, expired and expiring memories.

Read-only. The report only proposes candidates โ€” reviewing and acting on them (merge, supersede, extend, delete) is done via the regular memory tools. Similar pairs are always within a single workspace, ordered by similarity. Use the dream prompt for the full guided consolidation procedure.

Inputs: max_pairs, workspace, similarity_threshold. Returns: expired, expiring_soon, similar_pairs

get_memoryA
Read-only

Fetch full content of a single memory by name.

Use only when you already know the name โ€” e.g. after retrieve_memory() or recall_memory() which return names in their results alongside compact snippets. Do NOT use for search โ€” use retrieve_memory() for semantic/text search or recall_memory() for time-based queries like 'last week'. Unlike search, this also returns expired memories (expires_at in the past) โ€” check expires_at to tell; extend it via update_memory to bring one back.

Inputs: name (required), workspace. Returns: Full memory record returned by get_memory MCP tool.

list_memoriesA
Read-only

Browse all memories ordered by creation date (newest first). Returns full content (not snippets). Use to enumerate or audit without a specific query.

Inputs: tag, page, page_size, memory_type. Returns: page, items, total, page_size

list_workspacesA
Read-only

List all workspaces you are a member of (personal + shared).

Inputs: no named parameters. Returns: result

move_memoryA

Move a memory to a different to_workspace.

name: name of the memory to move. workspace: name of the target workspace (must be a member with write access). from_workspace: disambiguate source if the name exists in multiple workspaces.

Inputs: name (required), to_workspace (required), from_workspace. Returns: name, created

recall_memoryA
Read-only

Search memories by time expression + semantics. Returns names + metadata only.

Examples: "last week", "yesterday", "about Python last month". Use get_memory(name=...) to fetch full content of a specific result. Pass workspace= to search only within a specific workspace.

Inputs: query (required), n_results, workspace, memory_type. Returns: items

retrieve_memoryA
Read-only

Hybrid semantic + full-text search. Returns names + metadata only.

Use get_memory(name=...) to fetch full content of a specific result. Pass workspace= to search only within a specific workspace.

Inputs: query (required), limit, workspace, memory_type, +1 more. Returns: result

search_by_tagA
Read-only

Find memories by exact tag match. Returns all results (no pagination).

operation="AND" (default): memory must have ALL specified tags. operation="OR": memory must have AT LEAST ONE of the specified tags. Tags are case-insensitive. Use retrieve_memory() for semantic/text search or list_memories(tag=...) to browse a single tag with pagination.

Inputs: tags (required), operation. Returns: page, items, total, page_size

store_memoryA
Idempotent

Save a new memory. Idempotent: returns existing if content already stored.

name: human-readable name, unique within the workspace. workspace: name of the workspace to store into. Omit to store as a personal memory. force: skip near-duplicate check and store unconditionally. expires_at: optional UTC expiry; after it passes the memory is hidden from reads.

Inputs: name (required), content (required), memory_type (required), tags, +4 more. Returns: name, created

update_memoryB

Update an existing memory identified by name. Only provided fields are changed.

new_name: rename the memory to this name. workspace: disambiguate if the name exists in multiple workspaces. expires_at: set or extend the UTC expiry; omit to leave it unchanged.

Inputs: name (required), memory_type (required), tags, content, +4 more. Returns: name, created

Tool Schema Changelog

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

Loading tool schema changelogโ€ฆ

    TDQS

    A3.9/5.0

    Scored across 11 tools

    Disambiguation4/5

    Retrieval tools (retrieve_memory, recall_memory, get_memory, list_memories, search_by_tag) overlap in purpose, but descriptions clearly distinguish semantic/text search, time-based search, direct name lookup, browsing, and exact tag search. CRUD, workspace, and consolidation tools are clearly separate.

    Naming Consistency4/5

    Almost all tools follow a consistent snake_case verb_noun pattern (store_memory, update_memory, retrieve_memory, etc.). The one outlier is dream_report, which is noun-based rather than verb-based, but the set remains predictable overall.

    Tool Count5/5

    11 tools is well-scoped for a memory management server. Each tool appears to cover a distinct operation, and the count sits comfortably within the typical 3-15 range without obvious bloat.

    Completeness3/5

    Core memory lifecycle is covered: store, get, update, delete, move, search, list, and expiry handling. Notable gaps remain: there is no create/delete workspace tool despite list_workspaces, and no tool to add or remove tags despite search_by_tag, which could leave agents unable to set up tag-based searches.

    Maintenance

    ActivitySlowing
    ResponsivenessOutdated data

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      D
      maintenance
      Persistent memory and handoff intelligence layer for MCP agents. Most memory servers retrieve text โ€” Memory Nexus compounds operational context, learning from usage and progressively synthesizing observations into higher-order intelligence across sessions and tools.
      MIT
    • F
      license
      Not graded
      quality
      D
      maintenance
      Provides persistent AI agent memory using a local vector database for long-term semantic storage and short-term session scratchpads. It enables low-latency memory operations including search, storage, and bulk management without external cloud dependencies.
      -