✨ What is Eling?

Eling is a lightweight, unified second brain for AI agents. It merges 6 memory tiers via four MCP servers — and adds a Continuum Layer 8 orchestration tier that turns eling into a shared hub for multiple AI coding agents. The Blackbox Layer 2 flight recorder captures all agent telemetry (file reads, edits, shell commands, tool calls) and scores them with 11 context-efficiency metrics — turning raw observability into actionable optimization suggestions. The Obsidian Layer 6 gives agents local Markdown vault access — scoped project notes, daily logs, research notebooks, and memory review as plain files. No external databases, no cloud services needed for local operation (though it optionally syncs to Notion for online sharing and to Continuum for multi-agent orchestration).

Think of it as one memory stack that serves both the agent and the human — and one orchestration hub that serves every agent you run, with a flight recorder that watches everything they do:

📡 Layer 8: CONTINUUM — multi-agent orchestration hub (shared continuum.db, 15 continuum_* tools)
🧠 Layer 7: NOTION   — online brain, persistent, human-readable (optional)
📝 Layer 6: OBSIDIAN — local Markdown vault, project notes, daily logs, research (optional)
📚 Layer 5: KB       — FTS5 knowledge corpus for long-form knowledge
🕸️ Layer 4: CODE     — codegraph symbol intelligence
💎 Layer 3: FACTS    — SQLite + HRR + BM25 hybrid with trust scoring
🔎 Layer 2: BLACKBOX — flight recorder + telemetry + 11-metric efficiency scoring
⚡ Layer 1: BUILTIN  — MEMORY.md / USER.md (always-on, zero setup)

How the tiers work together

🧠 Notion as Online Memory (Layer 7)

Tier 7 (Notion) is what makes eling human-readable. While tiers 1–6 live locally as SQLite databases or flat files, Tier 7 optionally syncs high-trust facts to your Notion vault as permanent, well-formatted pages:

• eling_reflect <fact_id> — promote a high-trust fact to a Notion page
• eling_sync --direction push — batch-sync all high-trust facts to Notion
• eling_sync --direction pull — pull Notion pages back into the knowledge base
• eling_sync --direction all — bidirectional sync

This gives you a recoverable brain: even if your local SQLite databases are lost, your Notion vault retains the curated facts and permanent knowledge. The opinionated vault structure keeps the main page as a clean index of credential children, with all log entries auto-routed to a dedicated child page.

🚀 Quick Start

pip install eling

# Run the Notion-only MCP server (online/remote memory, 6 tools)
python3 -m eling mcp

# Run the local-layers MCP server (facts, KB, code, builtin, Blackbox, Obsidian, HRR — 38 tools)
python3 -m eling.as_brain.mcp_server
# or
eling as-brain

# Run the Blackbox flight recorder MCP (telemetry, scoring, baselines — 16 blackbox_* tools)
python3 -m eling blackbox mcp

# Run the Continuum Layer 8 orchestration hub (multi-agent, 15 continuum_* tools)
python3 -m eling continuum mcp
# or
eling-continuum

# Use the CLI
python3 -m eling --help

# If using OpenCode, install the lifecycle plugin:
eling-install-opencode

🔌 Agent Integration

One shared hub for every agent (Continuum Layer 8)

Continuum turns eling into a single MCP hub that all your coding agents connect to. Each agent gets isolated git worktrees, a shared orchestration registry, and two-tier knowledge (fundamental = binding rules, situational = semantic search) — all routed through eling's memory. Every entry is auto-attributed by the agent's MCP handshake name, so you always know which agent wrote what.

Wire all six agents in one command (Hermes, OpenCode, MiMo-Code, Zero, Claude Code, Codex):

# from the eling repo
chmod +x continuum/install.sh
continuum/install.sh --eling-home /shared/eling     # shared store for all agents
continuum/healthcheck.sh --eling-home /shared/eling # verify every agent is wired

Per-agent configs live in continuum/configs/; uninstall with continuum/uninstall.sh. See continuum/README.md for the full guide.

Non-tested agents connect via the stdio MCP servers — any MCP-compatible host can use eling (notion-only, 6 tools), as_brain (local layers, 33 tools), blackbox (flight recorder, 16 blackbox_* tools), and continuum (orchestration hub, 15 continuum_* tools).

🔎 Blackbox Flight Recorder (Layer 2)

Blackbox is eling's observability layer — a real-time flight recorder that captures everything your agents do and scores their efficiency. Think of it as a black-box data recorder for AI agents: it logs every tool call, file read, file edit, shell command, and subagent spawn, then runs 11 context-efficiency metrics on the collected data.

16 MCP tools for telemetry capture and analysis:

11 efficiency metrics (ported from Agent-Blackbox by Taewoo Park):

Agent support:

• Zero — auto-captures telemetry via the eling hook plugin; watch live streams with blackbox_watch_start
• Hermes — import past sessions with blackbox_ingest_hermes_session; live capture via as_brain MCP
• Any MCP agent — ingest events directly via blackbox_ingest

The Blackbox recorder feeds its findings into the Facts layer (Layer 3) for persistent causal memory, and baselines are stored per project across runs so you can track agent efficiency over time.

Hermes

Eling plugs into Hermes Agent at 3 levels:

1. MCP Server — add to ~/.hermes/config.yaml:

mcp_servers:
  eling:
    command: python3
    args: ["-m", "eling.mcp_server"]
    enabled: true

2. Memory Provider — sets default brain for remember/recall:

memory:
  provider: eling

3. Plugin — registers eling_remember + eling_recall as quick tools:

plugins:
  enabled:
    - eling
  eling:
    home: /root/.eling

OpenCode

Eling provides an OpenCode lifecycle plugin that auto-writes session memory:

# After installing eling, run this to install the plugin:
eling-install-opencode

# Or:
python3 -m eling install-opencode

This copies eling-memory.js to OpenCode's plugin directory and registers it in opencode.jsonc. The plugin hooks into:

• chat.message — stores user prompts as facts
• tool.execute.after — stores tool observations as facts
• event (session.idle / session.compacted) — pushes high-trust facts to Notion

The eling MCP server should also be configured in OpenCode (opencode.jsonc):

"mcp": {
  "eling": {
    "type": "local",
    "command": ["python3", "-m", "eling.mcp_server"],
    "enabled": true
  }
}

Zero

Eling provides a one-command installer for Zero terminal agent — hooks + skill + MCP in one go:

# After installing eling:
eling-install-zero
# Or:
python3 -m eling install-zero
# Preview without changes:
python3 -m eling install-zero --dry-run

This sets up:

Hook mapping

Manual config

If you prefer to wire it up yourself, add the MCP server to Zero's ~/.config/zero/config.json:

"mcp": {
  "eling": {
    "command": "python3",
    "args": ["-m", "eling.mcp_server"]
  }
}

📋 CLI Commands

python3 -m eling remember   "I learned that..."
python3 -m eling recall     "what did I learn about X"
python3 -m eling probe      "X"
python3 -m eling reason     ["X", "Y"]
python3 -m eling reflect    1                 # promote fact_id 1 to Notion
python3 -m eling verify                        # query verification status
python3 -m eling verify-spec                   # run spec-kit conformance

# Memory version control
python3 -m eling snapshot  --reason "pre_evolution"  # snapshot facts DB
python3 -m eling list-snapshots                       # list all snapshots
python3 -m eling rollback  <snapshot_id>              # restore to snapshot

# Zettelkasten linking + evolution
python3 -m eling link-stats                    # link graph stats
python3 -m eling linked-facts 1                # facts linked to fact_id 1
python3 -m eling evolve                        # merge near-duplicate facts
python3 -m eling stats
python3 -m eling export     --format markdown
python3 -m eling sync       --direction push   # facts → Notion

# Agent integration
python3 -m eling install-opencode              # install OpenCode lifecycle plugin
python3 -m eling install-zero                  # install Zero hooks + skill + MCP
python3 -m eling init-rules                    # write steering rules for AI agents

# Temporal search (v0.6.0)
python3 -m eling search-temporal "last 3 days" --category testing
python3 -m eling search-temporal "kemarin"     # Indonesian language support

# Per-fact versioning (v0.6.0)
python3 -m eling versioned-update 1 "Updated content" --reason "correction"
python3 -m eling version-history 1
python3 -m eling undo-to-version 1 --version-id 0
python3 -m eling versioning-stats

🌐 Notion Setup (Tier 5)

Optional — skip this if you only need local memory.

1. Create a Notion integration at https://www.notion.so/my-integrations

   • Give it a name (e.g. "Eling Brain")
   • Copy the Internal Integration Secret (starts with ntn_)

2. Share a parent page with your integration

   • Open the page you want as your second brain root
   • Click Share → Invite → select your integration
   • Copy the page URL and extract the page ID (the UUID in the URL, e.g. 38f7b66e-c7e0-813f-85b0-d37cef59c1f7)

3. Set environment variables:

export NOTION_API_KEY="ntn_..."
export NOTION_PARENT_PAGE_ID="38f7b66e-c7e0-813f-85b0-d37cef59c1f7"

Note-taking behavior

Once configured, eling auto-creates a 📋 Task Logs child page under your parent on first use:

📋 Hermes Vault (parent page — your configured root)
  ├── 📋 Task Logs        ← auto-created by eling
  │   ├── 💡 Eling test ← child pages from eling_reflect / remember(layer="notion")
  │   └── 💡 Another note
  ├── 🔑 API Keys...
  └── ...

Two ways to add notes to Notion:

Auto-routing by category

Content is automatically detected and routed to the right child page:

Example:

# Auto-routes to 🎯 Project Summaries
b.remember("Project done, deployed to production", layer="notion")
# Auto-routes to 🔑 Credentials
b.remember("DATABASE_URL = postgres://...", layer="notion")
# Auto-routes to 📋 Task Logs (no pattern match)
b.remember("General note", layer="notion")

All child pages under these category pages are full Notion pages — you can edit, move, share, or reference them normally.

Or pass them explicitly in code:

from eling.brain import Brain
b = Brain(
    notion_api_key="ntn_...",
    notion_parent_id="38f7b66e-..."
)
result = b.reflect(fact_id=1)
print(result)  # {"page_id": "...", "promoted": True}

# Or store directly as a note
result = b.remember("Quick note for Notion", layer="notion")
print(result)  # {"layer": "notion", "page_id": "...", ...}

Note: eling_reflect and remember(layer="notion") check availability at call time and return a clear error if any config is missing — no silent failures.

🕒 Temporal Search & Per-Fact Versioning (v0.6.0)

Eling v0.6.0 introduces time-aware fact retrieval and append-only per-fact versioning — never lose a piece of knowledge again.

Temporal Search

Query facts by time range using natural language — English or Indonesian:

python3 -m eling search-temporal "last 3 days"
python3 -m eling search-temporal "this week"
python3 -m eling search-temporal "kemarin"        # Indonesian: yesterday
python3 -m eling search-temporal "hari ini"       # today

Supported patterns:

Python API:

from eling.brain import Brain
b = Brain()

# Temporal search - English
results = b.search_temporal("last 3 days", category="testing")

# Indonesian
results = b.search_temporal("kemarin")

# All facts in a time window
results = b.search_temporal("", since_days=7)

Per-Fact Versioning

Every fact update is append-only — old versions are preserved in a fact_versions table:

# Update a fact — previous content is versioned
result = b.versioned_update(1, "Newer content", reason="corrected typo")
# → {"fact_id": 1, "version_id": 2, "previous": "Old content", "new": "Newer content"}

# Get version history
history = b.get_version_history(1)
# → [{"version_id": 0, "content": "Original...", "changed_at": "...", "reason": "initial"},
#     {"version_id": 1, "content": "Updated...", "changed_at": "...", "reason": "corrected typo"}]

# Undo to a specific version (also versioned!)
result = b.undo_to_version(1, version_id=0)
# → {"fact_id": 1, "version_id": 3, "restored_from": 0}

# Versioning stats
stats = b.versioning_stats()
# → {"versioned_facts": 42, "total_versions": 156, "version_operations": 114}

Available as MCP tools: eling_versioned_update, eling_get_version_history, eling_undo_to_version, eling_versioning_stats.

🧠 Memory Version Control

Eling provides Git-like snapshot and rollback for your facts database:

# Before destructive ops, create a snapshot
python3 -m eling snapshot --reason "pre_evolution"

# List available snapshots
python3 -m eling list-snapshots

# Rollback to a previous state (auto-backups current DB first)
python3 -m eling rollback 20260703-120000-123

Snapshots are file-level copies managed via snapshot.py. Available as MCP tools: eling_snapshot, eling_list_snapshots, eling_rollback.

🎯 Steering Rules

Teach your AI agent when to use eling's MCP tools. Auto-detects Cursor, Claude Code, OpenCode, Kiro, and Gemini:

cd your-project
python3 -m eling init-rules

This writes:

• Cursor: .cursor/rules/eling-memory-*.mdc
• Claude Code: .claude/rules/eling-memory-*.md
• OpenCode: Appends to AGENTS.md
• Generic: ELING_MEMORY.md in project root

Rules cover: when to store/retrieve memories, session lifecycle, and memory hygiene.

🔍 Vector Embeddings

Optional semantic search via sentence-transformers:

pip install eling[embeddings]
# or
pip install eling[all]

Enable when creating a Brain or set ELING_EMBEDDING_MODEL:

from eling.brain import Brain
b = Brain(embedding_model="all-MiniLM-L6-v2")

Hybrid search ranking: BM25 + Jaccard + HRR + cosine similarity from embeddings. Stored in a separate fact_embeddings table.

🛡️ Verify-on-Stop (Conditional + Universal)

Eling provides verify-on-stop nudges for AI agents that lack built-in verification (e.g., OpenCode, OpenClaw, Cursor, Windsurf). When running under Hermes, this feature automatically skips — because Hermes already has its own agent/verification_stop.py.

Universal mode — one shared brain for every agent

The as_brain MCP server can act as a universal brain for all connected agents at once. Set ELING_VERIFY_ALL_AGENTS=1 to force eling's verify-on-stop to stay active for every agent — including Hermes — so the shared brain provides verification regardless of harness. The default (unset) keeps the original behaviour: Hermes skips eling's nudges and relies on its own built-in verification.

# Run the shared brain with verification for ALL agents
ELING_VERIFY_ALL_AGENTS=1 python3 -m eling.as_brain.mcp_server

This powers multi-agent setups (e.g. Hermes + OpenCode + Zero sharing one as_brain instance) where you want a single source of truth for verification evidence.

How it works

1. Auto-detection — Eling detects the host agent from the MCP client's initialize handshake (clientInfo.name), which is more reliable than environment variable heuristics (prevents false Hermes detection when OpenCode runs under Hermes)
2. Agent auto-attribution — The handshake client name becomes the default source for brain_remember, so each agent's memories are tagged with its own identity without manual configuration (override with an explicit source argument)
3. File edit tracking — When code files are edited via hooks or MCP tools, eling records them in a verification ledger
4. Spec-kit conformance — If the project has spec-kit artifacts (specs/*/spec.md), eling checks whether code changes cover each spec requirement and includes gaps in the nudge
5. Verification nudge — If code was edited but no passing tests/verification was recorded, eling produces a [System: ...] nudge message
6. Recording — Agents can call brain_verify MCP tool (as_brain server) to record verification results (passed, failed, skipped)

Spec-kit Verification

Projects using spec-kit (Spec-Driven Development) get automatic spec conformance checking:

• Eling detects specs/<feature>/spec.md, plan.md, and tasks.md artifacts
• Requirements are extracted from spec markdown and matched against code files
• The eling_verify_spec tool returns coverage stats + uncovered requirements
• The standard eling_verify tool includes spec-kit results when spec_check=true
• Uncovered requirements are listed in the verification nudge for the agent to address

Usage via MCP

// Query current status
{ "method": "tools/call", "params": { "name": "eling_verify", "arguments": {} } }

// Record a passing verification
{ "method": "tools/call", "params": {
    "name": "eling_verify",
    "arguments": { "status": "passed", "command": "pytest", "output": "364 passed" }
} }

// Run spec-kit conformance check
{ "method": "tools/call", "params": {
    "name": "eling_verify_spec",
    "arguments": { "changed_files": ["src/main.py"] }
} }

// Combine both: verify + spec-kit
{ "method": "tools/call", "params": {
    "name": "eling_verify",
    "arguments": { "spec_check": true }
} }

Config

plugins:
  eling:
    adapter: auto       # auto-detect from env
    verify_on_stop: true

🏗️ Architecture

eling/
├── mcp_server.py              — JSON-RPC stdio server (Notion-only, 5 tools: eling_*)
├── as_brain/
│   └── mcp_server.py          — JSON-RPC stdio server (local brain + Blackbox, 33 tools)
├── blackbox/                   — Layer 2: Flight recorder & telemetry
│   ├── core.py                — TraceEvent, RunSummary, AgentMetadata
│   ├── store.py               — SQLite-backed event store
│   ├── score.py               — 11-metric efficiency scoring engine
│   ├── effectiveness.py       — Outcome scoring
│   ├── timeline.py            — Causal timeline builder
│   ├── mcp_server.py          — 16 blackbox_* MCP tools
│   ├── cli.py                 — Blackbox CLI subcommands
│   └── adapters/
│       ├── zero.py            — Zero stream-JSON adapter + plugin
│       └── hermes.py          — Hermes session DB adapter
├── continuum/                  — Layer 8: Multi-agent orchestration hub
│   ├── mcp_server.py          — JSON-RPC stdio server (15 continuum_* tools)
│   ├── store.py               — continuum.db: projects, agents, knowledge, plot, reservations
│   ├── worktree.py            — Isolated per-agent git worktree manager
│   ├── plot.py                — PLOT.md canonical protocol (unified-diff mutations)
│   └── continuum.sh           — Shared wrapper exec'd by every agent's MCP config
├── brain.py               — Orchestrator: routing + RRF fusion + sync + snapshot
├── config.py              — Layered config: env → json → defaults
├── hooks.py               — 15 lifecycle hooks + HookRegistry
├── verify_on_stop.py      — Verification ledger + nudge builder + spec-kit wiring
├── spec_kit.py            — Spec-kit artifact parser + coverage analyzer
├── snapshot.py            — Git-like snapshot & rollback for facts DB
├── rules.py               — Steering rules generator (Cursor, Claude Code, OpenCode)
├── privacy.py             — PII/secret stripping (19 patterns)
├── compress.py            — SHA-256 dedup + length compression
├── cli.py                 — `eling` CLI (18 subcommands, includes blackbox dispatch)
├── fact_memory_provider.py — Standalone facts layer provider (no Brain)
├── opencode_plugin/       — Bundled OpenCode lifecycle plugin
│   └── eling-memory.js
└── layers/
    ├── builtin.py         — Layer 1: MEMORY.md / USER.md loader
    ├── facts.py           — Layer 3: SQLite + HRR + BM25 + Embeddings + Trust + Zettelkasten + Temporal + Versioning
    ├── embeddings.py      — Optional vector embeddings (Mistral API + sentence-transformers)
    ├── hrr.py             — Holographic Reduced Representations (numpy)
    ├── code.py            — Layer 4: CodeLayer wrapper
    ├── code_index.py      — Pure-Python AST+regex code indexer
    ├── kb.py              — Layer 5: FTS5 + porter + trigram + RRF
    └── notion.py          — Layer 7: httpx Notion API client
    └── obsidian.py        — Layer 6: local Markdown vault client

⚡ Performance

• Lazy imports — numpy and httpx are imported only when their layer is first used, not at module load time
• import eling takes ~1.3s (was ~4.5s with module-level imports on Alpine)
• Pure-Python fallback when numpy unavailable (BM25-only retrieval still works)

📖 Documentation

• Architecture
• API Reference
• Migration Guide
• Hooks Reference

🤝 Credits

See CREDITS.md for full attribution — Hermes Agent (Nous Research), Agent-Blackbox / Taewoo Park, Continuum / Pouya Hasanamreji, spec-kit (GitHub), dusterbloom's holographic plugin, and all other contributors.

📜 License

MIT © 2026 PatrickNoFilter