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

v1.4.1 — field-store write/read hygiene (live handoffs, query packs, doctor smells) on top of the local field diary. 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. 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:

   sh

   Copy

   # 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:

   sh

   Copy

   # 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:

   sh

   Copy

   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.

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

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:

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:

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:

   sh

   Copy

   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.

Capture Reliability

Agent instruction files get an agent to read memory reliably at session start, but writes are less consistent — a long session can compress away the instructions, and some clients never load a rules file at all. Two mechanisms close that gap without depending on any agent's cooperation — commit, and the project brain updates itself; end a session without a journal, and the Stop hook blocks it.

Passive capture. With ai-memory init --install-hooks, every commit is recorded as episodic memory automatically:

ai-memory capture              # capture HEAD (this is what the post-commit hook runs)
ai-memory capture --commit abc1234

Each capture writes one file per commit at memory-store/episodic/commits/<date>/<short-hash>.md, with source: passive-capture and review_status: pending frontmatter — reviewable, and distinct from agent-written journal entries. It's idempotent per commit hash, needs no LLM, and is the raw material the next ai-memory dream consolidates or extracts facts from.

One file per commit rather than one shared file per day is what keeps the tree mergeable. Capture runs after the commit, so the record for commit N can only be committed by commit N+1 — it is always sitting uncommitted in between. A shared day file would therefore be a tracked, modified file at all times, and git refuses to merge into one (error: Your local changes to the following files would be overwritten by merge) before any merge machinery runs — including the semantic merge driver that would have resolved it. A distinct new file per commit is untracked instead, which git merges straight over, and two developers can never write the same path.

Noise commits (merges, [bot] authors, chore:/style:/ci:/build(deps) prefixes, lockfile-only changes) are skipped by default — audibly, with a skipped_reason and a --no-filter opt-out, never silently — and ai-memory dream --mode deep folds commit records older than 14 days into weekly summaries so episodic/commits/ never grows unbounded. See ROADMAP_CAPTURE_HOOKS.md for the full design.

Session-end enforcement. ai-memory session-start marks when a session began; ai-memory guard-journal exits non-zero (reason on stderr) if write_session_journal_tool hasn't been called since. ai-memory install --client <claude-code|gemini-cli|codex> --with-hooks wires this in automatically — SessionStart marks the session and injects a short context reminder, Stop blocks ending the session until the journal exists, and a pre-compaction event runs a non-blocking local consolidation checkpoint. ai-memory status reports local capture stats (last journal, commit captures, memories written in the last 7 days) so you can see whether capture is actually happening.

Proof, not just a claim. scripts/capture_rate_benchmark.py scripts a fully non-cooperative simulated agent — one that never calls write_session_journal_tool on its own — through 20 sessions in each mode:

Passive commit capture holds at 100% either way, since it runs off the git post-commit hook, not the client-side session hooks. Session journaling goes from 0% to 100% only once the Stop hook is wired in — a mechanism proof (the enforcement primitive cannot be silently skipped), not a statistical study of real-world agent compliance, which is separate, larger benchmark work (see ROADMAP.md Phase 5). Reproduce it yourself: python scripts/capture_rate_benchmark.py --sessions 20.

Failure memory. The highest-signal category for a coding agent — "I hit this exact error before, here's what fixed it":

ai-memory failure --error "KeyError: 'user_id' in handlers.py:12" --fix "Added a default value."

The error text is normalized (paths and line numbers stripped) into a stable signature, so a repeat of the same kind of error accumulates onto one growing memory-store/failures/<slug>.md entry — with an occurrences counter — instead of scattering into near-duplicate files.

Git-Native Trust

Two capabilities that only exist because memory lives in the same git repo as the code it describes — no vector-database or cloud memory product can offer either.

Semantic merge driver. Two branches that each append new facts to the same store file used to conflict on the shared last_updated line even when the actual content additions never overlapped. ai-memory init --merge-driver registers a custom driver that resolves the shapes team memory actually produces, and falls back to git's own textual merge for anything it can't safely resolve — never worse than not having it installed.

ai-memory init --merge-driver

What it merges, in order of preference:

This writes .ai-memory/**/*.md merge=memory-fabric to .gitattributes (committed and shared) and registers the driver command in local git config (per-clone by git's own design).

Teams: the per-clone half is the part that bites. .gitattributes travels with the repo, but the driver command does not — a teammate who clones and merges without registering it gets ordinary textual conflicts in .ai-memory/, with nothing to explain why. Two things close that gap: ai-memory doctor warns when a clone declares the driver but hasn't registered it, and any ai-memory init in such a clone registers it automatically.

Already mid-merge? You don't have to redo the merge — resolve-conflicts applies the same resolution to the conflict git already recorded, and stages what it resolves:

ai-memory resolve-conflicts
# resolved: .ai-memory/failures.md, .ai-memory/memory-store/episodic/2026-07-28.md
git commit

It exits non-zero and leaves the file untouched — conflict markers and all — for anything that needs a human, so it's safe to run on a whole conflicted merge.

Self-verifying citations. See the evidence field above — ai-memory verify is the command that checks citations still resolve and flags the ones that don't.

Global Memory

Developer-level preferences that apply across all projects are stored at:

global/directives.md is Tier 0: it is always loaded fully, bypassing token budgeting.

CLI Reference

ai-memory [--cwd <path>] [--json] [--debug-llm] <command>

Commands:
  init            Create .ai-memory/ scaffolding (--install-hooks, --merge-driver)
  status          Show memory status and capture stats
  doctor          Validate memory files and environment
  verify          Check evidence citations still resolve (--no-mark for a read-only report)
  failure         Record an error -> fix pair (--error, --fix, --tags)
  merge-driver    Git merge driver backend (invoked by git itself, not for direct use)
  resolve-conflicts  Resolve conflicted .ai-memory files of an in-progress merge (--all)
  capture         Record a commit as episodic memory (--commit, default HEAD)
  session-start   Mark session start (for client SessionStart hooks)
  guard-journal   Exit non-zero if no session journal was written (for client Stop hooks)
  install         Configure an MCP client to use memory-fabric (--client <name|all>)
  eval            Score memory quality or Dreaming quality
  bench           Coding-memory benchmark: memory-on vs memory-off retrieval
  dream           Run memory maintenance (--mode light|deep)
  migrate         Split legacy hand-written sections into store entries (--dry-run, --section, --no-llm)
  query           Search memory
  store           CRUD operations on semantic memory store
  sync-agents     Regenerate agent instruction files from canonical templates
  sync-global     Preview local-to-global promotions
  rollback        Restore from a snapshot (--to <name>, or --list to discover valid names)
  clean           Prune old snapshots/candidates (--keep-snapshots, --keep-candidates, --dry-run)
  diary           Local field diary (approve|revoke|mute|unmute|status|checkpoint|wipe) — off by default, never sent

Eval examples:

ai-memory eval
ai-memory eval --json
ai-memory eval --llm-review
ai-memory eval --dream latest
ai-memory eval --dream memory-20260601T140000_0400

Optional LLM review is never enabled by default. When requested, deterministic local scores remain the source of truth; LLM notes are secondary and inputs are sanitized before review.

ai-memory bench
ai-memory bench --json
ai-memory bench --fixture . --suite .ai-memory/evals/bench.yaml

bench scores whether later-session questions recover earlier decisions (memory-on vs an empty store). The built-in fixture and methodology live in benchmarks/coding-memory/. No LLM is required.

Built-in fixture (2026-08-15, ai-memory bench):

LLM Configuration & MCP Sampling

Memory Fabric uses Large Language Models (LLMs) to perform semantic memory consolidation (Dreaming) and qualitative evaluation. You can configure this in two ways: via direct environment variables or via zero-config MCP Sampling.

Resolution Precedence

When an operation requires an LLM, Memory Fabric resolves the provider in the following order:

1. Direct LLM Provider: Uses direct API calls if MEMORY_FABRIC_LLM_PROVIDER is set in the environment along with the corresponding API keys.
2. Native MCP Sampling: If no direct provider is set (or configured keys are missing) AND the command is executed within an active MCP session where the parent agent supports sampling, Memory Fabric delegates the LLM call to the agent itself.
3. Graceful Local Fallback: If neither is available, it degrades gracefully to local, deterministic regex-based deduplication without semantic synthesis.

Using MCP Sampling in Dreaming

MCP Sampling allows Memory Fabric to run semantic Dreaming and evaluations without configuring any local API keys or provider environment variables. Instead, it uses the host client's (e.g. Claude Code) native LLM connection.

How It Works

1. When your AI assistant calls the dream_tool or evaluate_memory_fabric_tool via MCP, the Memory Fabric server checks if the client session supports sampling capabilities.
2. If supported and no direct provider is configured, the server sends a create_message request back to the client.
3. The client assistant executes the LLM reasoning under the hood and returns the consolidated memory JSON or evaluation notes.

Key Benefits

• Zero-Config: No need to configure or expose GEMINI_API_KEY, OPENAI_API_KEY, or ANTHROPIC_API_KEY to the background MCP server process.
• Unified Context: The consolidation uses the same model and settings configured in your main coding assistant.

[!IMPORTANT] CLI Limitation: MCP Sampling relies on an active client-server session. Since the standalone terminal CLI (ai-memory) runs outside of an MCP client, running ai-memory dream from the command line cannot use MCP Sampling. To use LLM-based consolidation via the CLI, you must configure a direct LLM provider.

Client-Driven / Split-Tool Dreaming (Avoiding Deadlocks & Timeouts)

In some sequential/blocking MCP client environments (like IDE extensions or agentic loops), executing a nested MCP Sampling request (create_message) while the client is waiting for a tool response can lead to a JSON-RPC deadlock or execution timeouts.

To completely avoid this re-entrancy issue, Memory Fabric provides a client-driven split-tool dreaming protocol:

1. Prepare Payload: The agent/client first calls prepare_dream_payload_tool. This returns a consolidation_prompt and a candidate_store ID, but performs no blocking LLM reasoning.
2. Execute Locally: The client uses its own local LLM connection/context to execute the returned consolidation_prompt and captures the raw JSON output.
3. Apply Results: The client sends the raw LLM JSON response back by calling apply_dream_results_tool(candidate_store=..., llm_response=...). Memory Fabric parses the response, applies the consolidated changes to the candidate folder, and saves the new memory states.

Direct LLM Providers Configuration

To use the CLI with LLMs, or to bypass MCP Sampling with a specific provider, set the following environment variables:

Model quality matters for deep Dreaming. Consolidation quality — especially contradiction detection — degrades with small local models, and the failure is silent (the model simply returns an empty contradictions list). In our testing an 8B-class model (qwen3-vl:8b) missed a deliberately planted TTL conflict; prefer a ~14B+ model (or a hosted provider) when you rely on dream --mode deep for semantic conflict detection. As a safety net, Memory Fabric always runs a deterministic heuristic that flags store files with overlapping wording but divergent numbers, independent of the LLM's answer.

LLM Debugging

If you want to view the exact prompts sent and responses received by the LLM providers (Gemini, OpenAI, Anthropic, Ollama), you can enable LLM debug logging.

• Via the CLI: Pass the --debug-llm global flag before the subcommand:

  sh

  Copy

  ai-memory --debug-llm dream --mode deep
  

  This automatically prints logs to sys.stderr and appends them to a file named llm_debug.log (in .ai-memory/llm_debug.log if the directory exists, otherwise in the current working directory).

• Via Environment Variables: Set the MEMORY_FABRIC_LLM_DEBUG variable in your environment:

  • stderr: Output prompts and raw JSON responses only to sys.stderr.
  • 1 or true: Output to sys.stderr and write to the default llm_debug.log file.
  • /path/to/log.txt (or any other custom value): Append logs to a custom file path.

To protect your credentials and API keys, headers like Authorization and x-api-key are automatically redacted as [REDACTED] in the debug logs.

Write Safety

Every write path runs secret detection before saving. Detected secrets are replaced with [REDACTED_SECRET] and returned in the redactions field of the response. File locking prevents concurrent write corruption.

Eval scans existing memories for likely secrets but does not rewrite existing files. It reports what needs manual review.

Requirements

• Python >= 3.11
• mcp >= 1.0.0 (optional, required for MCP server only)
• rg (optional; ripgrep speeds up keyword search, Python fallback used when absent)

Contributing

Contributions are welcome. See CONTRIBUTING.md for setup, the checks CI enforces, and the store-first conventions. For anything security-sensitive, see SECURITY.md rather than opening a public issue.

License

MIT — © 2026 Memory Fabric Contributors. No telemetry, no account, no cloud.