Memória
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@Memória search my notes for ticket ABC-123 and its epic"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Agent Memory
English · Português · हिन्दी · עברית
Agent Memory keeps what Claude Code learns as Markdown files on your computer. It shows those notes as a graph and gives them back to Claude Code through an MCP server.
Until v0.4 the app was called Memória. Since v0.5 the names are in English too: the agent-memory command,
the agent-memory MCP server with the memory_search, memory_add and memory_related tools, and the ~/.agent-memory folder.
The old names still work (see Coming from v0.4).
Lessons, rules (do, don't) and task history.
Each note can be linked to a ticket (Jira, GitHub) and its epic.
Fast search (SQLite FTS5) by name, project, ticket, epic, tags and text.
A window with the graph, filters, a timeline, a form for new notes and Claude's MCP activity.
In English, Portuguese, Hindi and Hebrew.
Everything stays on your computer. Nothing is sent anywhere.
Install
No administrator rights needed.
macOS (Intel and Apple Silicon) and Linux, in Terminal:
curl -fsSL https://github.com/cx-guilherme-pinto/claude-memory-app/releases/download/v0.5.0/install.sh | shWindows 10/11, in PowerShell:
powershell -ExecutionPolicy Bypass -c "irm https://github.com/cx-guilherme-pinto/claude-memory-app/releases/download/v0.5.0/install.ps1 | iex"The installer:
installs uv (pinned version) if it's missing;
downloads the wheel and
constraints.txtand checks their SHA-256;installs the app with the uv-managed Python 3.12;
on Windows, installs the WebView2 Runtime (per user) if it's missing;
creates a shortcut:
~/Applications/Agent Memory.appon Mac, "Agent Memory" in the Start menu on Windows.
The commands go to ~/.local/bin (Mac/Linux) or %USERPROFILE%\.local\bin (Windows).
Portable version (no install)
Every release also has one zip per platform, with Python inside:
File | Open |
|
|
|
|
|
|
Unzip and open. No admin rights and no network needed. The workshop setup does this for you.
Connect it to Claude Code:
selftestprints the rightclaude mcp addcommand. On Mac,…/Memoria.app/Contents/MacOS/memoria selftest(the portable app keeps its old file name and shows as Agent Memory); on Windows,agent-memory.cmd selftest. After the one-line install, it's justagent-memory selftest.Prefer the one-line install command. A zip downloaded in a browser needs approval in System Settings → Privacy & Security: macOS blocks the app because it isn't signed. To unblock it in Terminal:
xattr -dr com.apple.quarantine ~/Applications/Memoria.app. On Windows, at "Windows protected your PC", click "More info" and then "Run anyway".Once it's connected to Claude Code, don't move the app.
The zips, the app folder and its launchers keep the
Memorianame: older portable copies can still update in place.
Details and decisions: docs/portatil.md.
Related MCP server: claude-brain
Connect to Claude Code
claude mcp add --scope user agent-memory -- "$HOME/.local/bin/agent-memory" mcpOn Windows: claude mcp add --scope user agent-memory -- "%USERPROFILE%\.local\bin\agent-memory.exe" mcp.
MCP tools (in Claude Code: mcp__agent-memory__memory_search and so on):
Tool | Purpose |
| Search notes. A ticket key ( |
| Save a note and return its path. |
| Linked notes, strongest relation first. |
type is lesson, do, dont or history.
Results use the same English names: {"total": 1, "results": [{"id", "path", "type", "title", "project", "task", "epic", "url", "tags", "files", "date", "cost_usd", "links", "excerpt", "score"}]}
(memory_related adds strength and reasons, for example {"link": 1, "task": 1}). Empty fields are left out.
Use
Command | What it does |
| Opens the window. |
| Opens the same interface in the browser. |
| Saves a note. |
| Searches from the command line. |
| Rebuilds the index from the Markdown files. |
| Installs the latest version. |
| Checks the installation. With |
| Shows or sets the language. |
| MCP server (used by Claude Code). |
In the window, click a node to see the note. Press / to search and Esc to close.
Search shows memories, projects and tasks/epics; pick one to focus it.
Filters choose the project, epic, task and period. The type list filters the map.
The timeline's play button shows the graph growing.
On Windows, if WebView2 is missing, the interface opens in the browser at 127.0.0.1.
The terminal prints the session address (http://127.0.0.1:<port>/?t=<key>).
The key is local and only works while the app is open.
Claude activity
The Claude activity button at the top of the column shows what Claude asked Agent Memory for over MCP:
four numbers: calls today, total calls, calls per minute (average of the last 10 minutes) and memories recalled (unique);
the last 100 recalled memories, with time, tool, title, project and task. A search with no results also shows up, with what was searched. Click a title to open the memory.
The view refreshes on its own. It isn't split by session.
The log lives in ~/.agent-memory/.index/activity.sqlite, on your computer, separate from the index
(agent-memory reindex keeps it). The searched text goes through the same secret redaction and is cut short.
Note contents are never logged. Old rows are dropped automatically; the total call count is kept.
Languages
The app speaks English, Portuguese, Hindi and Hebrew. By default it follows the system language:
macOS: the first supported language in System Settings › General › Language & Region;
Windows: the Windows display language;
others:
LANGUAGE,LC_ALL,LC_MESSAGESorLANG.
If none is supported, it uses English.
The picker at the top of the column switches the language at once and saves it in ~/.agent-memory/preferences.json.
From the command line: agent-memory language he. The AGENT_MEMORY_LANG variable overrides everything.
Hebrew opens right to left. Each note's text keeps its own direction.
Hindi and Hebrew use Noto Sans Devanagari and Noto Sans Hebrew, bundled with the app.
Technical terms stay in English in every language (MCP, token, ticket, tool names).
Where memories live
In ~/.agent-memory (or the folder set in AGENT_MEMORY_HOME):
~/.agent-memory/
<project>/<YYYY-MM-DD>-<slug>.md one note per file
config.json update URLs (optional)
.index/index-v3.sqlite index; safe to delete
.index/activity.sqlite MCP call log ("Claude activity" view)
preferences.json language chosen in the pickerEach note has a YAML header. The field names are part of the contract:
---
type: lesson # lesson | do | dont | history
title: Use the build cache
project: security-arena
task: ARENA-12 # optional: ticket key (or owner/repo#12)
epic: ARENA-1 # optional: epic or parent
url: https://company.atlassian.net/browse/ARENA-12 # optional
tags: [ci, cache] # optional
files: [orchestrator/plan.mjs]
date: '2026-10-16T13:25:33+01:00'
cost_usd: 1.25 # optional
links: [Avoid retries]
---
The build is 3x faster with the cache. See [[Avoid retries]].You can edit the files by hand. The app picks up the changes.
agent-memory reindex rebuilds the index at any time.
epic, url and tags came in v0.2 and are optional: v0.1 notes are still valid.
Tickets:
Jira keys are upper-cased (
abc-123→ABC-123); GitHub keys look likeowner/repo#12.Without
task, the app takes it fromurlwhen it's a Jira ticket or an issue/PR.A note without
epicinherits its task's epic, if another note names it.Searching
ABC-123finds that task's notes, its epic's notes and notes that mention it.
Link strength, strongest first:
explicit link ([[...]] or links), same task, same epic, same project, shared files.
Coming from v0.4
Nothing moves and nothing is rewritten. The old names keep working:
Until v0.4 | Since v0.5 |
|
|
|
|
MCP server | MCP server |
|
|
|
|
Header |
|
|
|
|
|
To use the new MCP name in Claude Code: claude mcp remove --scope user memoria, then the claude mcp add above.
Update
agent-memory updateFinds the latest version with
git ls-remote --tags(no API).Downloads the wheel and
constraints.txtand checksSHA256SUMS.Reinstalls with
uv tool install --force.Rebuilds the index. It never touches your notes.
On Windows, the install finishes in a new window. Close Claude Code first.
In the portable version, update downloads the new zip, checks it and swaps the app folder in place. The connection to Claude Code keeps working.
URLs come from ~/.agent-memory/config.json and, if missing, from the package's release.conf.
Keys read (an environment variable with the same name overrides them):
Key | Example |
|
|
|
|
|
|
| path to the |
Privacy
Notes, the index and the interface stay on your computer.
Before saving, the app redacts tokens, API keys (
sk-ant-…,ghp_…,AKIA…,xox…), JWTs, private keys and patterns likepassword=….The interface only listens on
127.0.0.1, with a random session key.The only network use is the version check (
git ls-remote) andagent-memory update. To turn it off:AGENT_MEMORY_NO_UPDATE=1.No embeddings and no models. Search is plain text (SQLite FTS5).
Troubleshooting
"This Python's SQLite has no FTS5": reinstall with uv's Python (
uv tool install --force --managed-python --python 3.12 <wheel>).Company proxy: uv and
agent-memory updateuseHTTPS_PROXYand the system certificates. For uv, also setUV_SYSTEM_CERTS=1.The window doesn't open:
agent-memory --browseropens the same interface in the browser."The search index of your memories is damaged": the app rebuilds the index from the notes by itself. If the message stays (on Windows, another process still has the file open), close Claude Code and the app, then run
agent-memory reindex. The notes are never touched."Warning: memory skipped": a note the app can't read (an error in the YAML header, not UTF-8, outside a project folder, a symlink to outside the folder). Fix the file: the app picks it up by itself.
Development
You need uv and Node.js 22 or newer.
uv sync
(cd web && npm ci && npm run build) && rm -rf src/memoria/web_dist && cp -R web/dist src/memoria/web_dist
uv run playwright install chromium
uv run pytestscripts/build.sh: UI, wheel, installers, portable zips andSHA256SUMSindist/.scripts/build_portatil.py: the portable zips (PORTATIL=0 scripts/build.shskips them).scripts/publish-release.sh: tag, release and assets. Calls to the hosting API live only increate_release()(Gitea or GitHub, fromRELEASE_PROVIDER).scripts/mudar-alojamento.sh: changes the hosting and the version in one go.scripts/lock.sh: regeneratesconstraints.txtand checks there are wheels for Mac and Windows.release.conf: the single source of URLs and versions.
Instructions for Claude: CLAUDE.md. Status, risks and how to publish: docs/HANDOVER.md. Versions follow SemVer. See the CHANGELOG.
License
MIT. See LICENSE.
The Bricolage Grotesque, Noto Sans Devanagari and Noto Sans Hebrew fonts ship with the interface
under the SIL Open Font License 1.1 (web/public/fonts/).
Available Tools
3 toolsmemory_addA
Save a memory as Markdown and return the file path. Secrets (tokens, keys, passwords) are redacted before writing.
Inputs: text (required), type (required), title (required), project (required), +7 more. Returns: result
memory_searchARead-only
Search memories by title, body, project, task, epic and tags. A ticket key (e.g. ABC-123) returns everything for that task and its epic. Filters: project, type, task and epic.
Inputs: epic, task, text, type, +2 more.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 3 tools
Each tool has a clearly distinct role: searching memories, finding related memories, and adding a memory. The boundaries are reinforced by descriptions that specify different inputs and outputs.
All tools share the `memory_` prefix and short suffixes, giving a predictable pattern. However, the suffixes mix action-oriented and relational terms (`search`, `add`, `related`), which is a minor inconsistency.
Three tools is a focused, minimal surface for a memory server and avoids bloat. It is slightly underprovisioned for full memory lifecycle management but appropriate for the likely intended lightweight use.
The server covers saving, searching, and relating memories, but lacks update and delete operations, leaving the lifecycle incomplete. Agents may hit dead ends when needing to modify or remove a memory.
Maintenance
Related MCP Connectors
Persistent memory layer for AI tools. Save and recall notes across Claude and other MCP clients.
Personal context for every AI: search, read, and write back to your private Markdown library.
Notes and memory assistant: one Markdown vault that Claude, ChatGPT and any MCP client can share.
Persistent memory for Claude Code and Cursor. Stop re-explaining your project every session.
Related MCP Servers
- AlicenseNot gradedqualityCmaintenanceEnables Claude Code to search and retrieve from a local knowledge base of markdown notes using hybrid semantic+keyword search, keeping data entirely offline.7 npmMIT
- AlicenseNot gradedqualityBmaintenanceEnables Claude Code to use local markdown notes as persistent, searchable memory with hybrid recall, episodic memory, note-graph traversal, design memory, and optional cloud sync.174 npmMIT
- AlicenseNot gradedqualityCmaintenanceEnables Claude Code to persist and recall cross-repository memories as Markdown files via save, search, read, and health-check tools, with local-only storage and no cloud services.9 npmMIT
- AlicenseNot gradedqualityCmaintenanceLets Claude read, search, create, update, and connect notes kept as plain local Markdown files, turning ideas, plans, and projects into an organized personal knowledge network. It can list and search notes, add new ones to the right area, change their status, and link related entries, without ever deleting them.73 npmMIT