Engram

Persistent memory for AI agents. In-process. No infra.

Give your AI agent the memory of a colleague who's worked with you for years — without cloud, API keys, or Docker.

⭐ Useful to you? Star it on GitHub — it's the simplest way to help others find Engram.

npm install -g @hbarefoot/engram
engram start

Your AI agent now has long-term memory. Two minutes, no setup, no cloud.

• 🧠 In-process — runs inside your agent's stack. No separate server to deploy, no IPC overhead, nothing to fork.
• 📴 Offline — local SQLite + bundled embeddings (~23 MB). No API keys, no data leaving your machine.
• 🔌 MCP-native — first-class Model Context Protocol integration with Claude Desktop, Claude Code, Cursor, Windsurf, and Cline.
• 🔐 Safety by default — automatic secret detection on every write. API keys, private keys, connection strings, JWTs blocked before they hit the database.

Why local-first, in numbers

Engram runs inside your agent's process — no service to deploy, no account, nothing leaving your machine. That design choice is measurable:

_{Measured on an Apple M4 Pro over 1,000 seeded memories — reproduce with npm run bench. These are footprint and latency numbers, not an accuracy claim: Engram doesn't try to out-rank Mem0 or Zep on memory benchmarks. The point is solid recall with none of the operational surface.}

Optional accuracy lift — still 100% local. If you already run a local model, the opt-in LLM layer sharpens fact extraction: entity-extraction accuracy climbs from 45.8% (rule-based) to 95.8% with the recommended henrybarefoot1987/engram-extract model (qwen3:1.7b) — +50 pts — without a single byte leaving your device.

Support Engram

Engram is free and MIT-licensed — and always will be. No paywalls, no tier-locked features, no telemetry. Every feature ships in the open-source package. Sponsorship is purely a way to fund continued development, not to unlock anything.

If Engram saves you time, you can sponsor it via Polar:

About Enterprise. Engram is MIT-licensed, so commercial use is already granted — you don't need to buy a license to use it at work. The Enterprise tier buys priority response on issues and dedicated help wiring Engram into your stack. For organizations whose policy precludes depending on MIT-licensed software, an optional commercial-license override is available on request. (Engram is maintained by a solo developer, so this is best-effort priority response, not a contractual SLA.)

Why Engram?

Most agent-memory products are services you run alongside your agent — Postgres, Docker, cloud accounts, API keys. Engram embeds inside your agent's process: a focused, stable npm package with practical guardrails.

Sources: @sunriselabs/lodis, Sunrise-Labs-Dot-AI/engrams, mem0.ai, github.com/getzep/zep, github.com/letta-ai/letta. See docs/competitive-intel.md for the full breakdown. Engram ships optional, on-device LLM extraction (v1.9+): point llm.* at a local model — the recommended henrybarefoot1987/engram-extract (Qwen3-1.7B, Apache-2.0) or any Ollama / OpenAI-compatible endpoint — to sharpen category/entity extraction (entity recognition +50 pts vs rules — 45.8% → 95.8% — with engram-extract (qwen3:1.7b) in our benchmark), still 100% local and off by default (the zero-config path stays rule-based, offline, and infra-free). Mem0/Zep/Letta build LLM extraction in via a cloud model; Lodis is LLM-free read/write with a broader feature surface — we list it honestly.

TL;DR — when each one fits. Pick Engram if you want a focused, stable, local-first memory layer with practical guardrails (secret detection, agent auto-discovery, desktop app), a simple 5-category mental model, and optional on-device LLM extraction when you want it. Pick Lodis if you want a knowledge-graph-style memory with 14 entity types and temporal supersession. Pick Mem0/Zep/Letta if you want cloud-LLM extraction built in and don't mind operating infrastructure for it.

Quickstart

1. Install

npm install -g @hbarefoot/engram

2. Start the server

engram start             # MCP + REST + Dashboard on localhost:3838
engram start --mcp-only  # MCP server only, stdio mode (for agent integration)

3. Connect your AI agent

Claude Code:

claude mcp add engram -- engram start --mcp-only

Claude Desktop — add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "engram": {
      "command": "engram",
      "args": ["start", "--mcp-only"]
    }
  }
}

Cline / Cursor / Windsurf — add the same mcpServers block to your editor's MCP config. The built-in dashboard at http://localhost:3838 has an Integration Wizard that auto-detects your installed agents and generates the config for you.

4. Use it

You:    "Remember that our API uses JWT tokens with 24-hour expiry."
Claude: (stores via engram_remember)

You:    (next day) "What authentication approach are we using?"
Claude: (recalls via engram_recall) — "JWT tokens, 24-hour expiry."

Memories persist across sessions, machine restarts, and even between different AI clients sharing the same Engram instance.

Memory that improves over time

Most memory systems are append-only stores: write once, retrieve forever, hope for the best. Engram learns.

• Feedback loop (engram_feedback) — when an agent recalls a memory, you or the agent can vote it helpful or unhelpful. Memories accumulate a score in [-1, 1]; consistently-unhelpful memories see their confidence decay automatically.
• Contradiction detection — when two memories conflict ("prefers Fastify" vs "switched to Express"), the consolidation engine flags them. The dashboard's Conflicts tab shows them side-by-side with four resolution actions: keep A, keep B, keep both, or dismiss.
• Deduplication on insert — identical memories (≥0.95 cosine similarity) are rejected. Near-duplicates (0.92–0.95) absorb the new content into the existing record. The store stays clean without manual pruning.
• Decay — memories that aren't recalled lose confidence over time and stop polluting future results.

The longer you use Engram, the sharper its recall gets.

MCP Tools

Engram exposes 6 tools to AI agents over stdio:

Memory categories

• fact — Objective truths about setup, architecture, or configuration.
• preference — User likes, dislikes, style choices.
• pattern — Recurring workflows and habits.
• decision — Choices made and the reasoning behind them.
• outcome — Results of actions taken.

Teach your agent to use Engram

Connecting the MCP server gives your agent the memory tools — but not the judgment to use them well. The bundled engram-memory skill is that judgment layer: it teaches an agent to recall at the start of a session, store durable decisions, corrections, and outcomes as they happen, and write results back at the end — without being told each time.

engram skill install                     # → ~/.claude/skills/engram-memory/
engram skill install --project           # → ./.claude/skills/  (commit it for your team)
engram skill install --platform agents   # → ~/.agents/skills/  (cross-framework)

Works in Claude Code, Claude Desktop, Cowork, or any framework that reads the Agent Skills spec (.agents/skills). The skill is vendored in the package, so it versions with Engram and updates land on the next engram skill install; engram skill uninstall removes it cleanly.

CLI Reference

engram start                       # Start MCP + REST + dashboard
engram start --mcp-only            # MCP server only (stdio mode)
engram start --port 3838           # Custom REST port

engram remember "<content>"        # Store a memory   (-c category -e entity -n namespace --confidence)
engram recall "<query>"            # Search memories  (-l limit -c category -n namespace --threshold)
engram forget <id>                 # Delete by ID
engram list                        # List memories    (-l limit --offset -c category -n namespace)
engram status                      # Health check

engram consolidate                 # Deduplicate, detect contradictions, decay
                                   # (--no-duplicates / --no-contradictions / --no-decay / --cleanup-stale)
engram conflicts                   # List unresolved contradictions
engram export-context              # Export curated context block
                                   # (-o file -f markdown|claude|txt|json -c categories --min-confidence ...)
engram import                      # Import from local sources
                                   # (-s cursorrules|claude|package|git|ssh|shell|obsidian|env --dry-run)

engram skill install               # Install the engram-memory agent skill
                                   # (--project → ./.claude, --platform agents → ~/.agents)
engram skill uninstall             # Remove the engram-memory skill

Run engram --help for the full flag list.

REST API

The REST API runs on http://localhost:3838 by default.

Web Dashboard

A built-in React dashboard at http://localhost:3838:

• Dashboard — Memory stats, recent activity, health gauge.
• Memories — Browse, filter, inline-edit, bulk-delete.
• Search — Semantic search with score breakdown.
• Statistics — Charts by category, namespace, and time.
• Health — Stale, never-recalled, low-feedback memories with one-click cleanup.
• Conflicts — Side-by-side contradiction resolution.
• Agents — Integration wizard that auto-detects installed AI clients and writes their MCP configs (with timestamped backups).
• Import — Wizard for cursorrules, .claude files, package.json, git config, SSH config, shell history, Obsidian, and .env.

How it works

1. Store: engram_remember runs content through secret detection, then embeds it locally using all-MiniLM-L6-v2 (~23 MB, CPU-only, downloaded once and cached at ~/.engram/models/). The embedding and metadata land in SQLite at ~/.engram/memory.db.
2. Recall: engram_recall embeds the query, fetches candidates via FTS5 + in-namespace embeddings, and scores them as (similarity × 0.45) + (recency × 0.15) + (confidence × 0.15) + (access × 0.05) + (feedback × 0.10) + fts_boost. Top results are returned and their access stats updated.
3. Deduplicate: on insert, identical memories (≥0.95 similarity) are rejected; near-duplicates (0.92–0.95) absorb new content into the existing row.
4. Learn: engram_feedback adjusts a memory's feedback_score and — after 5+ votes — bumps the confidence score up or down.
5. Protect: every write passes through pattern-based secret detection (OpenAI/Stripe/AWS/GitHub/Slack/Google keys, private keys, connection strings, JWTs, high-entropy strings). Detected secrets either reject the memory or redact the secret portion.

Configuration

Engram stores everything under ~/.engram/:

~/.engram/
├── memory.db          # SQLite database (memories + embeddings + FTS5 index)
├── config.json        # Server configuration
└── models/            # Cached embedding model

Defaults work out of the box. To customize:

{
  "port": 3838,
  "dataDir": "~/.engram",
  "defaults": {
    "namespace": "default",
    "recallLimit": 5,
    "confidenceThreshold": 0.3,
    "tokenBudget": 500,
    "maxRecallResults": 20
  },
  "embedding": {
    "provider": "local",
    "model": "Xenova/all-MiniLM-L6-v2"
  },
  "consolidation": {
    "enabled": true,
    "intervalHours": 24,
    "duplicateThreshold": 0.92,
    "decayEnabled": true
  },
  "security": {
    "secretDetection": true,
    "auditLog": false
  }
}

The llm.* block powers the optional local AI enhancement below. It is off by default (llm.provider: null); the zero-config path uses rule-based extraction and makes no LLM calls.

Optional: local AI enhancement (Ollama)

Engram works fully offline with zero AI dependencies. If you want a little more accuracy and already run a local model, you can optionally turn on "Layer 1" — and it stays 100% on your machine.

• Free, opt-in, off by default. Nothing changes unless you enable it.
• Local-first. Uses your own Ollama (default) or any OpenAI-compatible local server (LM Studio, llama.cpp). No cloud, no API key, no telemetry — your memory content never leaves your device.
• Graceful. Every call has a timeout and falls back to the built-in rule-based path if the model is slow, unreachable, or returns junk. Engram never crashes because a model is down.

What it improves when enabled: sharper category/entity/confidence on new memories, and an LLM confirmation step that reduces false-positive contradiction flags.

Recommended model: henrybarefoot1987/engram-extract. The layer's two jobs are classification, not generation — so a small model with constrained decoding (the model is forced to emit valid JSON) and thinking turned off is fast (sub-second), cool, and accurate. Pull it (or build it locally from the Modelfile):

ollama pull henrybarefoot1987/engram-extract
# …or build from source:
ollama create henrybarefoot1987/engram-extract -f models/engram-extract.Modelfile

Then set the model to henrybarefoot1987/engram-extract. It's a recommendation, not a lock-in — any Ollama or OpenAI-compatible model still works. See docs/llm/recommended-model.md for the base model, licensing, and how to pick the smallest model that beats rules on your hardware.

Attribution. henrybarefoot1987/engram-extract is built on Qwen3-1.7B (© Alibaba Cloud, Apache-2.0). Engram only adds the extraction prompt and the constrained-output configuration; the base model's weights, license, and notice are unchanged.

Enable it (desktop app): Preferences → AI Enhancement → toggle on, pick a model, Test connection, Save. The same tab shows a live status badge, activity stats (enhanced vs fallback extractions, contradictions filtered, average latency), and a recent-events list so you can see the layer actually working. Programmatically, GET /api/llm/status and GET /api/llm/stats expose the same data (all local — no telemetry).

Enable it (config file) — ~/.engram/config.json:

{
  "llm": {
    "provider": "ollama",
    "endpoint": "http://localhost:11434",
    "model": "llama3.2:3b",
    "apiKey": null
  }
}

First: ollama pull llama3.2:3b. Set "provider": null to turn it back off (the default). For an OpenAI-compatible local server, use "provider": "openai-compatible" and point endpoint at it (e.g. http://localhost:1234); apiKey is sent only if set.

Privacy note: "no memory data leaves your device" is only literally true when endpoint is local (localhost/127.0.0.1). If you point it at a non-local host, memory content is sent there for classification — the desktop AI Enhancement tab shows an explicit warning in that case. If the model is unreachable, a circuit breaker pauses the layer and Engram falls back to rule-based extraction with no added latency.

Advanced usage

Sandboxed evaluation

Redirect Engram's data directory to a throwaway location so it doesn't touch ~/.engram/memory.db. Useful for first-time evaluators, CI runs, or testing the desktop sidecar against a fresh DB:

# Via CLI flag (highest priority)
engram start --data-dir /tmp/engram-eval

# Or via env var
ENGRAM_DATA_DIR=/tmp/engram-eval engram start

# Works on every Engram command that touches the DB:
ENGRAM_DATA_DIR=/tmp/engram-eval engram remember "test memory" -c fact
ENGRAM_DATA_DIR=/tmp/engram-eval engram recall "test"
ENGRAM_DATA_DIR=/tmp/engram-eval engram status

Override priority: --data-dir flag > ENGRAM_DATA_DIR env var > dataDir in ~/.engram/config.json > default (~/.engram).

Namespace isolation

engram remember "Uses Next.js 14 app router" -n my-saas
engram remember "WordPress multisite + Redis" -n client-site

engram recall "what framework?" -n my-saas

Temporal queries

Time-range filtering is available via MCP and REST. Agents pass a time_filter object to engram_recall:

{
  "query": "deployment changes",
  "time_filter": { "after": "last week" }
}

{
  "query": "API decisions",
  "time_filter": { "after": "2026-01-01", "before": "2026-06-01" }
}

Supported shapes: after / before (ISO date or relative string like "3 days ago"), or period shorthand (today, yesterday, this_week, last_week, this_month, last_month, this_year, last_year).

Export context for documentation

engram export-context -f markdown -n my-project -o PROJECT_CONTEXT.md
engram export-context -f claude -o CLAUDE.md

Programmatic usage

Engram also works as a library inside your Node.js app:

import {
  loadConfig,
  getDatabasePath,
  getModelsPath,
  initDatabase,
  createMemory,
  recallMemories
} from '@hbarefoot/engram';

const config = loadConfig();
const db = initDatabase(getDatabasePath(config));

createMemory(db, {
  content: 'User prefers Fastify over Express',
  category: 'preference',
  confidence: 0.9
});

const results = await recallMemories(
  db,
  'preferred web framework',
  { limit: 5 },
  getModelsPath(config)
);

Contributing

See CONTRIBUTING.md for development setup, the versioning policy (npm + desktop bump together), and the release checklist. The project's licensing and sustainability stance is in BUSINESS_MODEL.md — short version: pure OSS, MIT forever, no paywalls.

git clone https://github.com/HBarefoot/engram.git
cd engram
npm install
npm run dev

If Engram is useful to you, here's how to help:

• ⭐ Star the repo — the loudest signal that this is worth continuing.
• 🐛 Open an issue — bug, feature request, or "we use Engram at <company> for <thing>" stories all welcome.
• 💬 Start a discussion — design questions, integration ideas, "how would I…" — all good.
• 💜 Support Engram — sponsor via Polar to fund continued development. No tier-locked features; sponsorship goes straight to keeping the project shipping.

Feedback

Using Engram? Tell me what's working and what isn't — open a Discussion, file feedback, or run engram feedback from the CLI. No telemetry, ever — Engram never phones home, so the only feedback I get is what you choose to send.

Find Engram on Glama

Engram is listed in the Glama MCP directory and the official MCP Registry as io.github.HBarefoot/engram.

License

MIT © 2026 HBarefoot