Memory Fabric MCPサーバー

Memory Fabric

File-first, local-first memory layer for MCP-compatible AI coding assistants.

Memory Fabric gives AI tools like Claude Code, Cursor, and GitHub Copilot a consistent, project-aware context layer without locking you into one model, editor, cloud provider, or operating system.

Memory is stored as human-readable Markdown with YAML frontmatter. No vector database. No cloud account. No embeddings required.

---

Features

• MCP-native: exposes memory tools through the standard Model Context Protocol
• File-first: Markdown files are the source of truth, inspectable and commit-ready
• Local-first: core reads and writes work offline
• Store-first: facts live in memory-store/, one per file; root maps (architecture.md, decisions.md, ...) are generated views rebuilt by Dreaming, never hand-written
• Captures itself: every git commit is recorded as episodic memory automatically — no agent cooperation required — via an opt-in post-commit hook
• Git-native merge: an optional custom merge driver lets two branches' memory merge as cleanly as their code, instead of conflicting on a shared timestamp line
• Self-verifying: memories can cite the file/line/commit they depend on; ai-memory verify flags citations that rotted
• Learns from failure: write_failure_memory_tool deduplicates repeat occurrences of the same error into one growing record instead of scattering near-duplicates
• Secret-safe: API keys and credentials are redacted before writing
• Token-budget aware: assembles context within limits; never slices files mid-document
• Quality eval: scores memory usefulness and Dreaming before/after results locally
• Unicode-safe: works with any human language
• Graceful degradation: works without rg, git hooks, or Dreaming configured

---

Status

v1.4.5 — field diary pack/analyze and live posture pinning. Live on PyPI. Core CLI and MCP tools work end-to-end. See ROADMAP.md for what shipped, what's in progress, and what's next.

---

Demo

The CLI and self-capture flow: initialize a project, write a memory as a plain Markdown file in your repo, and watch a git commit record itself as episodic memory with no agent cooperation. The cross-tool moment — an agent writes memory in one tool and a different tool reads it back — is storyboarded in DEMO.md for the full video.

---

Privacy — no telemetry

Memory Fabric never sends data anywhere. There is no account, no cloud, no analytics, no phone-home. The core read and write paths make no network calls at all; the only optional network requests are a PyPI version-drift check and an LLM-provider preflight, both of which you can turn off with --offline. Your memory is plain Markdown in your own git history — it never leaves your machine unless you push it.

An optional, off-by-default field diary can record operational counts on your machine after you run ai-memory diary approve. That is not the session journal, and it is not telemetry: nothing is transmitted. MEMORY_FABRIC_DIARY=0 is a kill switch. ai-memory diary pack builds a shareable folder under the app data field-packs/ directory (Windows: %APPDATA%/memory-fabric/field-packs/); you copy that folder yourself. ai-memory diary analyze scores product signals from a pack or the local diary. This is a deliberate guarantee, not a default we might change.

---

Installation

From PyPI (recommended)

Requires Python ≥ 3.11.

pipx install memory-fabric          # CLI only
pipx install "memory-fabric[mcp]"   # CLI + MCP server

Or with plain pip (inside a virtual environment):

pip install memory-fabric          # CLI only
pip install "memory-fabric[mcp]"   # CLI + MCP server

Zero-install one-off run (requires uv):

uvx --from "memory-fabric[mcp]" memory-fabric-mcp   # starts the MCP server on stdio

From GitHub (latest, pre-release)

# CLI only
pipx install "git+https://github.com/elViRafa/agentic-memory.git"

# CLI + MCP server
pipx install "memory-fabric[mcp] @ git+https://github.com/elViRafa/agentic-memory.git"

Or with plain pip (inside a virtual environment):

pip install "git+https://github.com/elViRafa/agentic-memory.git"          # CLI only
pip install "memory-fabric[mcp] @ git+https://github.com/elViRafa/agentic-memory.git"  # + MCP

Upgrading

Because Memory Fabric is in active development, we recommend upgrading regularly:

1. Upgrade the Package:

   # If installed via pipx:
   pipx upgrade memory-fabric
   
   # If installed in a virtual environment via pip:
   pip install --upgrade "memory-fabric[mcp]"
   
   # If installed from GitHub (pre-release):
   pipx install --force "memory-fabric[mcp] @ git+https://github.com/elViRafa/agentic-memory.git"
   

2. Refresh the MCP client config (uvx installs): If your client config launches the server via uvx (the default written by older ai-memory install runs), uv serves the build it cached the first time and never re-resolves an unpinned spec — restarting the client is not enough and can leave the server several releases behind without any signal. After upgrading, either:

   # Re-write the client config (now pins the exact installed version, or points
   # at the local memory-fabric-mcp binary when one sits next to the CLI):
   ai-memory install --client <your-client> --project
   
   # ...or drop the stale cached build so uvx re-resolves:
   uv cache clean memory-fabric
   

   ai-memory doctor warns when the local version drifts from the latest on PyPI or when a different ai-memory installation shadows this one on PATH.

3. Restart MCP Clients: After upgrading, restart your IDE (Cursor, VS Code) or assistant process (Claude Code) to ensure the client reloads the updated memory-fabric-mcp server.

4. Refresh Local Projects (Optional): If you have projects initialized with older versions of Memory Fabric, navigate to the project directory and run:

   ai-memory init --install-hooks
   

   This will safely refresh starter templates and local git hook integration to the latest format without overwriting your existing memory markdown files.

Windows terminals: the CLI emits UTF-8 (memory content legitimately contains em-dashes and bullets). Windows Terminal and PowerShell 7 render it out of the box; on legacy consoles (PowerShell 5.1 with an OEM code page) set chcp 65001 or PYTHONUTF8=1 if you see replacement characters.

Or clone and install in editable mode for local development:

git clone https://github.com/elViRafa/agentic-memory.git
cd agentic-memory
pip install -e .          # CLI only
pip install -e ".[mcp]"   # CLI + MCP server

---

Quick Start

1. Initialize a project

ai-memory init

Creates .ai-memory/ in the current directory with starter sections and a .gitignore.

2. Check health

ai-memory doctor

3. Evaluate memory quality

ai-memory eval
ai-memory eval --json

eval scores whether memories are useful for coding assistants. It checks section coverage, starter-template content, summary quality, metadata, retrieval readiness, and likely secrets.

If .ai-memory/ exists, reports are saved under ignored local files:

.ai-memory/evals/latest.json
.ai-memory/evals/latest.md
.ai-memory/evals/<timestamp>-memory.json
.ai-memory/evals/<timestamp>-memory.md

If .ai-memory/ does not exist yet, eval prints a pre-init report only and creates no files.

4. Query memory

ai-memory query "authentication"

5. Run maintenance (Dreaming)

ai-memory dream --mode light
ai-memory dream --mode deep

Dreaming creates a snapshot before maintenance. You can evaluate whether a Dreaming run improved memory quality:

ai-memory eval --dream latest
ai-memory dream --mode light --eval

Dream eval compares the pre-dream snapshot to current memory and reports score delta, changed files, improvements, and regressions.

---

MCP Server — Install

One command per client. Safe to re-run: merges with your existing config, never overwrites it, and backs up the original file if it can't be parsed.

One-click install

These call uvx --from "memory-fabric[mcp]" memory-fabric-mcp under the hood — same canonical invocation as ai-memory install, just without the CLI step. Requires uv.

Client	Command
Claude Code	ai-memory install --client claude-code
Claude Desktop	ai-memory install --client claude-desktop
VS Code	ai-memory install --client vscode
Cursor	ai-memory install --client cursor
Windsurf	ai-memory install --client windsurf
Codex	ai-memory install --client codex
Antigravity	ai-memory install --client antigravity
Gemini CLI	ai-memory install --client gemini-cli
Cline	ai-memory install --client cline
All detected clients	ai-memory install --client all

Add --project to write project-scoped config instead of the global/user config (where the client supports it). Add --dry-run to preview the change as a unified diff without writing anything. Add --uninstall to remove only the memory-fabric entry, leaving everything else in the file untouched.

Available MCP Tools

Tool	Description
initialize_memory_fabric_tool	Create .ai-memory/ scaffolding in a project
read_combined_context_tool	Load Tier 0 directives + prioritized memory within token budget
read_section_tool	Read a single memory section by name
keyword_search_tool	Search memory with ripgrep or Python fallback
write_local_memory_tool	Update steering sections (framework-rules, ubiquitous-language) only. Writing facts or root maps is rejected (store-first) — root maps are generated from memory-store/; use write_memory_store_tool
propose_memory_patch_tool	Preview proposed memory changes without applying
dream_tool	Run memory maintenance / consolidation (--mode light
prepare_dream_payload_tool	Prepare candidate snapshot and return consolidation prompt for client-driven dreaming
apply_dream_results_tool	Apply LLM consolidated JSON output to candidate snapshot and save changes
evaluate_memory_fabric_tool	Evaluate local memory quality
evaluate_dream_quality_tool	Evaluate a Dreaming run against a snapshot
write_memory_store_tool	Write a memory file to a semantic store path (e.g. architecture/decisions/auth); accepts an optional evidence citation list
read_memory_store_tool	Read a single memory-store file by its semantic path
list_memory_store_tool	List files in the memory store, optionally filtered by prefix/tags
delete_memory_store_tool	Remove a memory-store file by its semantic path
write_session_journal_tool	Append a timestamped session journal entry (episodic memory)
write_failure_memory_tool	Record an error → fix pair; repeat occurrences of the same error accumulate onto one entry

---

Agentic Architecture (making agents use Memory Fabric automatically)

Registering the MCP server is necessary but not sufficient — AI agents will not use its tools unless explicitly instructed to do so.

Memory Fabric provides an automated Agentic Architecture to handle this for you.

When you run ai-memory init in your project, the CLI automatically deploys a complete suite of agent instruction files tailored for every major AI tool:

Target	File(s) Created
Gemini CLI / Codex / Antigravity	AGENTS.md
Grok (TUI / Build / Agent harness)	AGENTS.md (primary project rules); full docs + MCP registration in ~/.grok/config.toml or .mcp.json; see also .grok/docs/user-guide/13-memory-fabric.md when installed for the client
Cursor IDE	.cursor/rules/memory-fabric.mdc
Windsurf IDE	.windsurf/rules/memory-fabric.md
Cline / Generic IDE Agents	.agents/rules/memory-store.md, .agents/rules/dreaming.md
Claude Code	CLAUDE.md (created or appended)
GitHub Copilot	.github/copilot-instructions.md (created or appended)

Single Source of Truth (Syncing)

All these files are generated from two canonical content blocks inside the Memory Fabric Python package. Running ai-memory sync-agents regenerates every platform file from these templates, guaranteeing perfect consistency. If you used ai-memory init --install-hooks, a pre-commit git hook runs this sync automatically on every commit.

In the shared user files (AGENTS.md, CLAUDE.md, .github/copilot-instructions.md) the fabric content lives inside managed marker blocks (<!-- >>> memory-fabric:… >>> --> … <!-- <<< memory-fabric:… <<< -->); sync only ever rewrites content between the markers, so anything you add outside them is preserved byte-for-byte.

Project Directives (v1.1)

Beyond the package's own protocol instructions, you can define project directives — hand-curated development guidelines shared by every AI agent working on the repo. A directive is simply a root section file .ai-memory/<name>.md with role: steering frontmatter (like the built-ins framework-rules and ubiquitous-language), plus two optional boolean routing keys:

Key	Meaning	Default
sync	include the directive in the per-tool files generated by sync-agents	true
context	inject the directive as Tier 0 into read_combined_context	true for the built-ins; false for user directives (file-reading agents already get the synced copy)

ai-memory sync-agents composes every sync: true steering section into one project-directives document and distributes it as .agents/rules/project-directives.md, .cursor/rules/project-directives.mdc, .windsurf/rules/project-directives.md, and a managed marker block in AGENTS.md/CLAUDE.md/.github/copilot-instructions.md — so AGENTS.md-only readers (Codex, OpenCode, Gemini) receive them too. Deleting the last synced directive tears all of that down again.

For CI, ai-memory sync-agents --check is a drift gate: it writes nothing and exits 1 when any generated file is out of date. ai-memory doctor additionally warns when the always-loaded context surface exceeds its token budget (default 3000, MEMORY_FABRIC_DIRECTIVE_BUDGET), when a hand-edited steering file contains something that looks like a secret, and when a stale pre-1.1 fabric block sits outside the managed markers.

For full details on the architecture, see AGENTIC_ARCHITECTURE.md.

Why this matters

Without these instruction files, agents will often explain they "don't use MCP tools automatically," or worse, they will write memory markdown files directly using native file-system tools—bypassing secret scanning, token budgeting, and background Dreaming management.

By simply running ai-memory init, you get zero-configuration, plug-and-play capability across the entire agent ecosystem.

---

Using Memory Fabric in Other Projects

You can use a single installed instance of Memory Fabric across all your coding projects:

1. Scaffold the target project: Navigate to your target project's root folder and initialize it:

   cd /path/to/other-project
   ai-memory init --install-hooks
   

   This automatically creates .ai-memory/, deploys the Agentic Architecture rule files, sets up .gitignore, and installs the git post-commit hooks.

2. Global MCP Configuration: Verify your AI assistant's configuration points to your globally installed memory-fabric-mcp executable. The MCP tools automatically parse and use the active project's path passed by the assistant as the cwd argument, enabling a single global MCP server registration to support all your local workspaces.

---

Project Memory Layout

Memory Fabric is store-first: memory-store/ is the only place agents write facts by hand, one fact per file. The root map files (architecture.md, decisions.md, debt.md, schemas.md, index.md) are generated views rebuilt by Dreaming from their matching memory-store/<category>/ subtree — never edit them directly, they carry generated: true frontmatter and a hand edit gets folded back into the store for review on the next Dream. The exception is the steering tier: any root section with role: steering frontmatter (the built-ins framework-rules.md and ubiquitous-language.md, plus any project directives you create) is hand-curated — humans and agents may edit these files directly (review via MR). The built-ins are always loaded into context in full, never generated and never evicted by the token budget; user directives default to context: false and reach agents through the sync-agents per-tool files instead.

.ai-memory/
|-- index.md                  # generated discovery index
|-- architecture.md           # generated map of memory-store/architecture/
|-- schemas.md                # generated map of memory-store/schemas/
|-- decisions.md              # generated map of memory-store/decisions/
|-- debt.md                   # generated map of memory-store/debt/
|-- ubiquitous-language.md    # hand-curated steering directive (always loaded)
|-- framework-rules.md        # hand-curated steering directive (always loaded)
|-- memory-store/             # the source of truth — one fact per file
|   |-- index.md              # generated store-wide index
|   |-- architecture/...
|   |-- decisions/...
|   |-- debt/...
|   |-- schemas/...
|   |-- failures/<slug>.md    # error -> fix pairs, deduplicated by normalized signature
|   |-- episodic/<date>.md    # agent-written session journals
|   `-- episodic/commits/<date>/<short-hash>.md  # one passively captured commit each
|-- evals/       # ignored local quality reports
|-- snapshots/   # ignored rollback baselines
|-- private/     # ignored personal notes + session markers
`-- .gitignore

All memory files use YAML frontmatter:

---
store_path: architecture/decisions/auth-service
summary: "One-line fallback used when the file exceeds the token budget."
priority: high
tags: [api, auth]
schema_version: "1.3"
last_updated: 2026-06-01T12:00:00-04:00
evidence: [src/auth.py:42, "commit:abc1234"]
---

Your memory content here.

The optional evidence field lets a memory cite what it depends on — a file, a file:line, or a commit:<hash>. Run ai-memory verify to check those citations still resolve; a memory citing a file that was renamed or deleted gets flagged instead of quietly rotting.

---