Memory Fort
GalaxyRuler/memory-fort
Memória persistente local e privada para agentes de IA cross-tool
Ver no GitHubServidor MCP que fornece memória persistente compartilhada entre ferramentas de IA, mantendo todos os dados localmente em arquivos markdown. Integra-se automaticamente com Claude Code, Codex, ChatGPT e outras plataformas, permitindo que agentes de IA retenham contexto entre sessões sem depender de APIs externas ou bancos de dados.
Instalação
npx convez-ai GalaxyRuler/memory-fort1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Manter contexto histórico em sessões de programação com IA
- Criar uma base de conhecimento versionável em git para decisões de projetos
- Recuperar informações relevantes em conversas com múltiplas ferramentas de IA
- Construir um grafo de conhecimento privado integrado com Obsidian
Veredito

Solução bem pensada para o problema real de fragmentação de memória entre ferramentas de IA. Arquitetura sólida (markdown + BM25 + busca semântica) com foco em privacidade e portabilidade. Sem imposições de vendor lock-in — a integração com Obsidian é elegante.
README
A tradução para o português deste README ainda está na fila.
Memory Fort gives every AI coding session a shared long-term memory: observations flow in automatically from Claude Code, Codex, Antigravity, Hermes, Pi, ChatGPT, and OpenCode. MCP integrations, including OpenClaw in v1, can log and recall memory on demand; search results include detailed provenance receipts explaining ranking decisions; a curated wiki of markdown pages grows over time; and retrieval (BM25 + semantic + graph) surfaces the right context at session start. No database. No external service. No API key to get started.
Your memory is a folder of plain text files — a git repo, an Obsidian vault, and a typed knowledge graph all at once.
Why Memory Fort?
Memory Fort does not require a cloud account, a running database, or a paid API key to get started; its local vault and lexical search run from plain markdown files.
- Your data, your machine. Everything lives under
~/.memory/as markdown files you can read, edit, grep, and version-control. - No vendor lock-in. Open schema, plain text format, vault is just a git repo.
- No account required to start. Lexical search (BM25 + graph) works on day one with zero API keys.
- Obsidian-native. Open
~/.memory/in Obsidian and get a knowledge graph, backlinks, and full-text search for free. - Cross-tool hooks. Claude Code, Codex, Antigravity, Hermes, Pi, ChatGPT, and OpenCode write to the same vault automatically.
- MCP-only clients. OpenClaw uses the same vault through MCP without passive capture in v1.
Quickstart
npx memory-fort init
Interactive wizard asks ≤4 questions (all pre-defaulted), detects your installed tools, and wires everything. Press Enter to accept all defaults.
Prerequisites: Node.js ≥ 20. Nothing else. No Docker, no database, no API key.
# Search immediately (no key needed)
memory-fort grep "your query"
# Browse and search in the UI
memory-fort dashboard
How it works
System architecture
%%{init: {
'theme': 'base',
'themeVariables': {
'primaryColor': '#1e293b',
'primaryTextColor': '#f8fafc',
'primaryBorderColor': '#475569',
'lineColor': '#38bdf8',
'secondaryColor': '#0f172a',
'tertiaryColor': '#1e1b4b'
}
}}%%
flowchart TD
subgraph Tools ["🤖 AI Tools"]
CC["Claude Code"]
CX["Codex"]
AG["Antigravity"]
HE["Hermes"]
PI["Pi"]
OC["OpenClaw / CovenCave"]
CD["Claude Desktop"]
VS["VS Code"]
GPT["ChatGPT"]
OCODE["OpenCode"]
end
subgraph Capture ["📥 Capture"]
H["⚡ Hook Scripts<br/>(PreToolUse · PostToolUse · UserPromptSubmit · Stop)"]
MCP_IN["🔌 MCP Server<br/>(log_observation)"]
end
subgraph Vault ["💾 ~/.memory/ (vault)"]
RAW["📝 raw/YYYY-MM-DD/<br/>session.md"]
WIKI["📚 wiki/<br/>projects · decisions · lessons<br/>tools · threads · procedures"]
IDX["🗂️ index.md · schema.md<br/>preferences.md"]
EMB["🧠 embeddings/<br/>*.jsonl sidecars"]
end
subgraph Curation ["🧹 Curation"]
COMPILE["⚙️ memory compile<br/>(distills raw → wiki)"]
end
subgraph Retrieval ["🔍 Retrieval (three routes)"]
SS["① Session-start hook<br/>(automatic push)"]
SEARCH["② MCP memory.search<br/>(BM25 + vector + graph + RRF)"]
GREP["③ memory grep<br/>(ripgrep over raw + wiki)"]
end
CC & CX & HE & PI & OCODE -->|hooks fire| H
AG & CD & VS & OC & GPT & OCODE -->|MCP calls| MCP_IN
H --> RAW
MCP_IN --> RAW
RAW --> COMPILE
COMPILE --> WIKI
WIKI --> IDX
IDX --> SS
WIKI & RAW & EMB --> SEARCH
RAW & WIKI --> GREP
SS -->|context injected| Tools
SEARCH -->|ranked results| Tools
GREP -->|matching lines| Tools
classDef tool fill:#0284c7,stroke:#0ea5e9,stroke-width:2px,color:#fff;
classDef capture fill:#0f766e,stroke:#14b8a6,stroke-width:2px,color:#fff;
classDef vault fill:#4c1d95,stroke:#7c3aed,stroke-width:2px,color:#fff;
classDef curation fill:#b45309,stroke:#f59e0b,stroke-width:2px,color:#fff;
classDef retrieval fill:#9d174d,stroke:#db2777,stroke-width:2px,color:#fff;
class CC,CX,AG,HE,PI,OC,CD,VS,GPT,OCODE tool;
class H,MCP_IN capture;
class RAW,WIKI,IDX,EMB vault;
class COMPILE curation;
class SS,SEARCH,GREP retrieval;
How memories reach your AI tools
%%{init: {
'theme': 'base',
'themeVariables': {
'primaryColor': '#1e293b',
'primaryTextColor': '#f8fafc',
'primaryBorderColor': '#475569',
'lineColor': '#38bdf8',
'secondaryColor': '#0f172a',
'tertiaryColor': '#1e1b4b'
}
}}%%
flowchart LR
subgraph V ["💾 ~/.memory/ (vault)"]
S["📄 schema.md"]
I["📄 index.md"]
L["📄 log.md"]
P["⚙️ wiki/preferences.md"]
W["📂 wiki/projects/<cwd>.md"]
WK["📚 wiki/ pages"]
EM["🧠 embeddings/"]
R["📝 raw/"]
end
subgraph Routes ["🔄 Retrieval routes"]
R1["(1) Session-start<br/>automatic push<br/>every session"]
R2["(2) MCP search<br/>on-demand pull<br/>BM25 + vectors + graph"]
R3["(3) grep<br/>CLI / manual"]
end
subgraph AI ["🤖 AI Tool Context"]
CTX["⚡ Session context<br/>(prompt prefix)"]
TOOL["🔌 Agent tool call<br/>(MCP response)"]
CLI2["💻 CLI output"]
end
S & I & L & P & W -->|reads on session start| R1
WK & EM & R -->|hybrid retrieval + RRF| R2
R & WK -->|ripgrep| R3
R1 -->|stdout injected by tool| CTX
R2 -->|ranked snippets| TOOL
R3 -->|matching lines| CLI2
classDef vault fill:#4c1d95,stroke:#7c3aed,stroke-width:2px,color:#fff;
classDef route fill:#0f766e,stroke:#14b8a6,stroke-width:2px,color:#fff;
classDef ai fill:#0284c7,stroke:#0ea5e9,stroke-width:2px,color:#fff;
class S,I,L,P,W,WK,EM,R vault;
class R1,R2,R3 route;
class CTX,TOOL,CLI2 ai;
Route (1) fires automatically — you always get your top context injected. Routes (2) and (3) are on-demand (agent or human asks).
Supported tools
memory-fort install claude-code # Claude Code (full hooks + plugin)
memory-fort install codex # Codex desktop + CLI (hooks + MCP)
memory-fort install antigravity # Google Antigravity / Gemini (MCP + live-capture plugin)
memory-fort install hermes # Hermes agent (YAML hooks + MCP in ~/.hermes/config.yaml)
memory-fort install pi # Pi coding agent (YAML hooks in ~/.pi/config.yaml)
memory-fort install chatgpt # ChatGPT (HTTP/SSE bridge for ChatGPT Connectors)
memory-fort install openclaw # OpenClaw (MCP server in ~/.openclaw/openclaw.json)
memory-fort install opencoven # OpenCoven / CovenCave (read-only daemon readiness check)
memory-fort install opencode # OpenCode (MCP config + selected event plugin)
memory-fort install claude-desktop # Claude Desktop (MCP only)
memory-fort install vscode # VS Code (MCP only)
See docs/cli.md for full CLI reference (all subcommands, flags, examples).
All installs are non-destructive and idempotent — sentinel-block writes, re-running is safe. The OpenCoven / CovenCave target is read-only: it checks the coven CLI and the local coven.daemon.v1 health contract, but does not launch sessions or write Memory Fort config. It is part of the OpenCoven family.
OpenClaw support is MCP-only in v1: the installer preserves/updates its MCP config, but it does not install passive capture hooks or automatic observation capture.
OpenCode support features MCP config wiring and selected event plugin installation for hook capture, with CLI diagnostics integration.
ChatGPT connects through an HTTP/SSE bridge server (default localhost:3100) that translates MCP stdio into HTTP Server-Sent Events, which ChatGPT's "Connectors" feature can consume. You can manage the bridge via the CLI (memory-fort chatgpt-bridge start|stop|status) or configure the port in ~/.memory/config.yaml under chatgpt.bridge_port. Autostart on Windows is supported via the HKCU Run registry key (best-effort).
# Undo any integration cleanly
memory-fort uninstall claude-code
memory-fort uninstall chatgpt
memory-fort disconnect --all
Retrieval modes
| Mode | Needs | When to use |
|---|---|---|
| Lexical (default) | Nothing | Day 1, offline, private projects |
| Voyage embeddings | VOYAGE_API_KEY | Hosted semantic retrieval |
| OpenAI embeddings | OPENAI_API_KEY | Alternative to Voyage |
| Ollama (local) | Ollama running locally | Full local, no cloud at all |
Switch any time: edit ~/.memory/config.yaml or re-run memory-fort init.
Search results include provenance receipts showing BM25, embedding, and graph signals — visible in the dashboard and MCP API.
When Voyage embeddings are enabled, Memory Fort uses contextualized text embedding: each wiki page is embedded with a graph-topology header (path, type, relations, tags, backlinks) prepended, improving recall for graph-aware queries. Enable via retrieval.embeddings.contextualized: true in config.yaml.
Evidence posture
Memory Fort avoids reproduced-score and third-party benchmark claims unless they have been reproduced locally. Vendor-reported benchmark numbers must stay labeled as vendor-reported, and public claims should point to release evidence or implemented local behavior.
Current local evidence is intentionally narrower than a benchmark claim:
| Area | Current public claim | Evidence status |
|---|---|---|
| Default storage | Markdown + YAML files under ~/.memory/ | Implemented local package behavior |
| Default search | Lexical search works without an API key | Implemented local package behavior |
| Search provenance | Search results include detailed provenance receipts showing ranking decisions | Implemented in MCP response (provenance field) and dashboard UI |
| Optional retrieval | Semantic and graph-assisted retrieval are available when configured | Local smoke evidence is recorded in docs/release-evidence/2026-06-06-v1.1-credibility.md |
| Package surface | Package uses the files whitelist in package.json | Local npm pack --dry-run --json evidence is recorded in the release evidence note |
Memory Fort does not currently publish a reproduced LongMemEval score or a reproduced third-party benchmark row. Use vendor benchmark numbers only as vendor-reported claims, not as Memory Fort certification.
Wiki schema
Memory Fort organizes curated knowledge by entity type:
| Type | Directory | Purpose |
|---|---|---|
projects | wiki/projects/ | Codebases and work efforts |
decisions | wiki/decisions/ | Architecture and tooling choices, with alternatives |
lessons | wiki/lessons/ | Reusable facts learned from incidents |
references | wiki/references/ | Papers, posts, talks |
tools | wiki/tools/ | Libraries and services |
threads | wiki/threads/ | Narrative arcs across a stretch of work |
procedures | wiki/procedures/ | Reusable step-by-step workflows |
Pages link via typed graph edges (uses, depends_on, supersedes, contradicts, caused_by, fixed_by, derived_from). Plain YAML frontmatter — no database required.
Dashboard
memory-fort dashboard
# → http://127.0.0.1:4410/memory/
Built-in React dashboard:
- Browse the wiki, search (BM25 + semantic + graph), review proposed pages, inspect graph health metrics.
- Audit ranking decisions using search provenance receipts (expandable detail showing BM25, embedding, and graph weights).
- Manage integrations using client toggles to enable or disable individual clients (configured via
clients.*map inconfig.yaml, with verify checks automatically skipping disabled clients). - Securely manage API credentials via API key management (masked secrets fields stored outside the vault with test-then-save validation flow).
- Check and manage the ChatGPT bridge status (verify checks:
chatgpt.bridge.runningandchatgpt.bridge.mcp). - Review lifecycle proposals (
wiki/compile-proposed/) — dispute and supersede candidates staged for human review before any wiki edit.
Roadmap
- Optional SQLite-FTS index — rebuildable cache for sub-10ms lexical search at large vault sizes
- Community integrations — pull requests welcome; hook pattern documented in
docs/architecture.md
License
Memory Fort is open source under the GNU General Public License v3.0 (GPL-3.0-only).
Commercial use is permitted under the GPLv3 terms. There is no separate commercial license requirement for this repository.
For contributors / private dev repo
After cloning, install the pre-push gate:
npm run install:dev-hooks
This gates every git push origin through scan:leaks so personal tokens can never accidentally reach the public repo.