Noshy — Persistent Memory for AI Agents

ICM-compatible. MCP-native. Works with any LLM.

Noshy gives your AI agent real memory — not note-taking, not context stuffing, not a vector database you have to manage. Store facts, search across sessions, build knowledge graphs. It's what ICM wanted to be, re-built to work everywhere.

                     Noshy
          ┌───────────┼───────────┐
          │   MEMORIES            │   MEMOIRS
          │   (time-bound)        │   (permanent)
          │                       │
          │  ┌───┐ ┌───┐ ┌───┐    │   ┌───┐
          │  │bug│ │fix│ │pref│    │   │doc│
          │  └───┘ └───┘ └───┘    │   └───┘
          │    │       │     │     │
          │    └───┬───┘     │     │
          │  ┌─────┴─────┐   │     │
          │  │   GRAPH   │   │     │
          │  │  relations │   │     │
          │  └───────────┘   │     │
          └──────────────────┴─────┘
                    │
          ┌─────────┴──────────┐
          │   HYBRID SEARCH    │
          │  keyword+semantic  │
          │       +graph       │
          └────────────────────┘

Why Noshy

• LLM-powered extraction — not regex. Uses any OpenAI-compatible API to extract structured facts from transcripts
• Hybrid search — keyword + semantic + graph recall in one query
• ICM compatible — import your existing ICM databases, uses the same schema
• MCP native — works with Claude Code, Hermes, Codex, Copilot, and any MCP client
• Any embedding provider — OpenAI, fastembed (local, free), or Hermes API server
• Zero dependencies — core runs on Python stdlib. fastembed and OpenAI are optional
• Single binary feel — one Python file does everything
• TOML config — ~/.noshy/config.toml with env var overrides, no code changes needed
• Graceful shutdown — SIGTERM/SIGINT handlers WAL-checkpoint the database and close connections cleanly
• Retry with backoff — LLM extraction retries on 429/5xx with exponential backoff (3 attempts)
• Input validation — empty topics, summaries, and titles are rejected before hitting the database
• Session hooks — automatic memory extraction at session end, no manual calls required
• Numpy-vectorized cosine — cluster detection and semantic search are ~50x faster with numpy (optional, pure-Python fallback included)
• Rotating request logs — set NOSHY_LOG_FILE or run in a container to get ~/.noshy/noshy.log (5MB x 3 rotation)
• Contradiction detection — scans memory pairs in a similarity band and asks the LLM whether they conflict; persists confirmed pairs as contradicts edges so future recalls warn about them
• Async extraction queue — hand off long transcripts without blocking on the LLM; a periodic sweep drains the queue in the background

Hermes vs Noshy

Hermes has built-in memory that persists across sessions, but it's small and limited to keyword matching. Noshy replaces that with a proper database that understands what you meant, not just what words you used.

What Noshy adds that Hermes doesn't have by default:

• Semantic search — finds memories by meaning, not just keyword matching. Ask about "that proxy issue" and it finds the entry about SOCKS5 configuration even if the word "proxy" isn't in it.
• Unbounded storage — Hermes caps memory at ~3,500 chars. Noshy stores unlimited entries in SQLite.
• Richer memory types — memories (facts/decisions), memoirs (permanent knowledge), and concepts (related ideas linked together). Hermes just has flat text entries.
• Graph relationships — links related memories so recalling one pulls up connected ones.
• Contradiction awareness — Noshy detects when two memories conflict (e.g. opposite preferences, replaced decisions) and surfaces warnings on recall so agents don't act on stale info.
• Session context injection — automatically surfaces what you were working on, recent decisions, and active projects when a session starts.
• Decision timeline — chronological log of what was decided and why, so you can trace back.
• Pattern detection — notices when the same solution keeps coming up and suggests turning it into a skill.
• Weight decay — older, less-referenced memories fade in relevance automatically instead of cluttering results.

Quick Start

# Install from PyPI (recommended)
pip install noshy

# Start the HTTP server + dashboard
noshy serve
# → http://127.0.0.1:8720/

# Or run as an MCP stdio server
noshy mcp

Or install from source:

curl -fsSL https://raw.githubusercontent.com/noshkoto/Noshy/main/install.sh | sh

Usage

CLI

# Store a memory (optional --ttl, --importance auto, --project)
python3 server.py store "deploy-config" "Deploy uses Cloudflare Pages with GitHub Actions"

# Recall (add --json for machine output)
python3 server.py recall "deployment config"

# List projects with counts and last activity
python3 server.py projects

# Delete: by id, by topic, or wipe an entire project
python3 server.py delete --id 01J...
python3 server.py delete --topic "old-bug" --scope onboarding
python3 server.py delete --project staging --yes

# Maintenance
python3 server.py purge                  # delete expired
python3 server.py consolidate-clusters   # merge near-duplicates
python3 server.py find-contradictions    # detect + link conflicting memory pairs
python3 server.py sweep                  # purge + decay + consolidate + drain queue

# Async extraction queue (hand off long transcripts without blocking on the LLM)
python3 server.py queue --file session.txt --session-id sess-2026-06-18
python3 server.py process-queue --limit 10

# Import from ICM
python3 server.py import /path/to/icm/memories.db

# Stats
python3 server.py stats

MCP Server (Claude Code, Hermes, Codex, Copilot)

Add to your MCP client config:

Claude Code (~/.claude/mcp_servers.json):

{
  "mcpServers": {
    "noshy": {
      "command": "python3",
      "args": ["/path/to/noshy/server.py", "mcp"],
      "env": {
        "NOSHY_EMBED_PROVIDER": "openai",
        "OPENAI_API_KEY": "sk-..."
      }
    }
  }
}

Hermes (config.yaml):

mcp_servers:
  noshy:
    command: "python3"
    args: ["/path/to/noshy/server.py", "mcp"]
    env:
      NOSHY_EMBED_PROVIDER: "openai"
      OPENAI_API_KEY: "sk-..."

Codex CLI (~/.codex/mcp.json):

{
  "mcpServers": {
    "noshy": {
      "command": "python3",
      "args": ["/path/to/noshy/server.py", "mcp"]
    }
  }
}

MCP Tools

HTTP API

# Store
curl -X POST http://127.0.0.1:8720/tools/call \
  -H 'Content-Type: application/json' \
  -d '{"name":"noshy_store_memory","arguments":{"topic":"my-topic","summary":"What to remember"}}'

# Recall
curl -X POST http://127.0.0.1:8720/tools/call \
  -H 'Content-Type: application/json' \
  -d '{"name":"noshy_recall","arguments":{"query":"search keywords"}}'

# Stats
curl http://127.0.0.1:8720/stats

# Recent memories (JSON)
curl 'http://127.0.0.1:8720/memories?limit=25'

Web Dashboard

The HTTP server also serves a zero-dependency web dashboard. Start the server
and open the root URL in a browser:

python3 server.py http
# then visit http://127.0.0.1:8720/

Dashboard features:

• Hyrule color palette — Rupee emerald, Navi fairy blue, Sunset amber accents on a twilight dark background
• Animated gradient orbs — floating radial gradients with CSS grid background art
• Project picker — custom dropdown with gradient-tinted selection, animated chevron, count pills, click-outside/Esc to close
• Hybrid search — keyword + semantic + graph in one query box; memoirs included
• Cluster view — surface groups of near-duplicate memories and merge them in one click
• Inline delete — hover a card, click the trash icon to remove it (with custom confirm dialog)
• Dark / light theme — auto-detected, manually toggleable, persisted to localStorage
• Animated stat counters — live database stats with skeleton loaders on first paint
• Toast notifications — feedback on store/delete/consolidate actions
• Pagination — ?page=1&limit=25 for large databases

When NOSHY_HTTP_TOKEN is set, the dashboard shows a token prompt modal on
first load. The token persists in localStorage across reloads. A "Forget"
button clears it. API routes enforce auth; / and /health stay public.

Session Hooks

Noshy can automatically extract memories when a session ends. Drop the hook
into your Hermes workflow or call it from any MCP client:

from hooks import on_session_end

result = on_session_end(transcript, project="my-project", max_memories=8)
# → {"extracted": 5, "ids": [...], "concepts": ["deploy", "ci"]}

The hook skips transcripts shorter than 100 characters and returns structured
results with extracted memory IDs and discovered concepts.

Python API

For scripts and apps, Noshy ships a small Python API with decorators that
make any function self-remembering:

import noshy

@noshy.remember(topic="deploy", importance="high")
def deploy(env):
    return f"deployed to {env}"

deploy("prod")                  # auto-stores: deploy -> 'deployed to prod'

# Scope memories to a project (and inherit tags) for a block of code
with noshy.session(project="checkout-bugfix", tags=["sprint-23"]):
    do_work()                   # every @remember inside picks up the project

noshy.recall("deploy")          # hybrid search returns matching memories

Useful keyword arguments on @noshy.remember:

• importance="auto" — let the LLM classify each memory (critical/high/medium/low)
• on_error=True (default) — exceptions are stored as high-importance memories
• capture_args=True — include arg names in the summary; arguments whose
  names look like secrets (password, token, api_key, ...) are auto-redacted
• skip_if=lambda r: r is None — don't store certain return values
• ttl_seconds=... — auto-expire after N seconds

For long-running sessions, noshy.extractor.stream_extract(chunks) yields
memories incrementally as transcript chunks arrive.

Embedding Providers

Noshy auto-detects the best available embedding provider. Set NOSHY_EMBED_PROVIDER to override:

# With OpenAI
export OPENAI_API_KEY="sk-..."
python3 server.py http

# With free local embeddings
pip install fastembed
python3 server.py http

# Keyword-only (no embeddings)
NOSHY_EMBED_PROVIDER=none python3 server.py http

Platform Setup

macOS

# Install Python 3.10+ if needed
brew install [email protected]

# Install Noshy
curl -fsSL https://raw.githubusercontent.com/noshkoto/Noshy/main/install.sh | sh

# Optional: local embeddings
pip3 install fastembed

Linux

sudo apt install python3   # Debian/Ubuntu
sudo dnf install python3    # Fedora

curl -fsSL https://raw.githubusercontent.com/noshkoto/Noshy/main/install.sh | sh

Windows

# Install Python from python.org (check "Add to PATH")

# Download Noshy
Invoke-WebRequest -Uri https://github.com/noshkoto/Noshy/archive/refs/heads/main.zip -OutFile noshy.zip
Expand-Archive noshy.zip -DestinationPath $env:USERPROFILE\.noshy
Rename-Item $env:USERPROFILE\.noshy\Noshy-main $env:USERPROFILE\.noshy\src

# Run
python $env:USERPROFILE\.noshy\src\server.py http

Docker

# Build the image from the included Dockerfile
docker build -t noshy .

# Run it (data persists in a named volume)
docker run -d --name noshy \
  -p 8720:8720 \
  -v noshy-data:/data \
  -e OPENAI_API_KEY=sk-... \
  noshy

# Or with HTTP auth enabled
docker run -d --name noshy \
  -p 8720:8720 \
  -v noshy-data:/data \
  -e NOSHY_HTTP_TOKEN=$(openssl rand -hex 32) \
  noshy

# Optional build flags
docker build --build-arg WITH_FASTEMBED=1 -t noshy .   # bake local embeddings
docker build --build-arg WITH_SQLITE_VEC=0 -t noshy .  # skip the vec extension

The image runs as a non-root user, exposes a /health endpoint, and uses
/data as a persistent volume.

HTTP authentication

By default the HTTP server is unauthenticated and binds to 127.0.0.1 only.
To expose it on a network or behind a proxy, set a bearer token:

export NOSHY_HTTP_TOKEN="$(openssl rand -hex 32)"
python3 server.py serve --host 0.0.0.0

Clients must then send Authorization: Bearer <token> on every request. The
/health endpoint and the dashboard HTML at / stay public so probes and
human visitors still work.

Configuration

Noshy supports two configuration methods. Environment variables always take
precedence over the config file.

Config file

Create ~/.noshy/config.toml (or set NOSHY_CONFIG to a custom path):

[noshy]
db-path = "~/.noshy/memories.db"
embed-provider = "openai"
embed-model = ""
api-base = "http://127.0.0.1:8642/v1"
model = "hermes-agent"
http-host = "127.0.0.1"
http-port = 8720

Environment variables

Database migrations

Noshy auto-migrates the database schema on startup (v1 through v4). No manual
steps required. New columns are added transparently; existing data is preserved.

Architecture

┌─────────────────────────────────────────┐
│              Noshy MCP Server            │
│  ┌──────────┐ ┌────────┐ ┌───────────┐  │
│  │Extractor │ │ Store  │ │  Embedder │  │
│  │(LLM API) │ │(SQLite)│ │(OpenAI/   │  │
│  │ + retry  │ │ +migrate│ │ fastembed) │  │
│  │          │ │ factory│ │ +numpy cos│  │
│  └──────────┘ └────────┘ └───────────┘  │
│         │          │           │         │
│         └──────────┼───────────┘         │
│                    │                     │
│        ┌───────────┴────────┐            │
│        │   Hybrid Search    │            │
│        │ keyword semantic   │            │
│        │      + graph       │            │
│        └────────────────────┘            │
│                    │                     │
│           ┌────────┴───────┐             │
│           │  MCP / HTTP    │             │
│           │  (stdio+API)   │             │
│           └────────────────┘             │
│                    │                     │
│           ┌────────┴───────┐             │
│           │  Session Hooks │             │
│           │  (auto-extract)│             │
│           └────────────────┘             │
└─────────────────────────────────────────┘

Import from ICM

# Import memories from an existing ICM database
python3 server.py import ~/.config/icm/memories.db

# Verify
python3 server.py stats

The schema is compatible — memories, memoirs, concepts, and metadata all transfer. Graph edges and feedback are preserved when available.

Comparison

Roadmap

• Web dashboard
• Semantic search over memoirs (auto-embedded on store)
• Automatic, importance-aware memory decay
• Consolidation that prunes merged duplicates
• Python decorator for automatic function memory (@noshy.remember)
• Memory importance prediction (importance="auto")
• Streaming extraction (extractor.stream_extract)
• Project isolation (list_projects / delete_project)
• Graph-based memory consolidation (find_clusters / consolidate_clusters)
• HTTP bearer-token auth (NOSHY_HTTP_TOKEN)
• Real Dockerfile (multi-stage, non-root, healthcheck)
• Integration test suite (pytest tests/)
• PyPI release (pip install noshy)
• Streaming extraction MCP tool (noshy_stream_extract)
• Dashboard polish (project picker, cluster view, inline delete, theme toggle)
• TOML config file (~/.noshy/config.toml)
• Schema auto-migration (v1-v4)
• Graceful shutdown (SIGTERM/SIGINT + WAL checkpoint)
• LLM extraction retry with exponential backoff
• Session context, decision timeline, and pattern detection MCP tools
• Session-end auto-extraction hooks
• Glassmorphism dashboard redesign
• Hyrule color palette (emerald, fairy blue, sunset amber)
• Numpy-vectorized cosine similarity (~50x faster)
• Dashboard extracted to dashboard.html (server.py 1832 -> 1012 lines)
• Dashboard auth UI (token prompt modal, localStorage, Forget button)
• Rotating HTTP request logs (NOSHY_LOG_FILE)
• Input validation and silent-except cleanup
• Contradiction detection (noshy_find_contradictions + contradicts edges on recall)
• Async extraction queue (noshy_queue_extraction / noshy_process_queue)
• 56 tests (was 47)
• Per-user database isolation (multi-tenant — one DB per token)

License

Apache 2.0 — same as ICM. Built as a drop-in improvement.

"Your agent shouldn't forget what you fixed last week."