mnemos

The autopilot knowledge base for your coding agent.

Install once. From then on, your agent builds itself a structured knowledge base of your project — while you code. No prompts to remember, no remember() calls, no API to learn.

Single Go binary. Embedded SQLite. Zero cloud. No Docker. No Python. No Node runtime.

Agent (Claude Code / Cursor / Kiro / Gemini CLI / Codex / ...)
    ↓ MCP stdio
mnemos serve
    ↓
Auto-compiled knowledge base (~/.mnemos/mnemos.db)

What makes mnemos different

Every memory server stores text. Mnemos compiles a knowledge base.

While other servers expect you (or a carefully-tuned prompt) to decide when to store and when to retrieve, mnemos runs a full pipeline in the background:

Agent action → mnemos auto-pipeline:
                ├── Quality gate        (reject/rewrite low-value content)
                ├── 3-tier dedup        (hash → fuzzy → semantic)
                ├── Auto-summarize      (extractive, fast; LLM if available)
                ├── File linking        (extract identifiers, link to code)
                ├── Type classification (episodic / long_term / semantic / working)
                ├── Quality scoring     (for retrieval ranking)
                └── Decay scheduling    (so knowledge base stays relevant)

Retrieval:
                ├── Hybrid search       (FTS5 + optional semantic + RRF)
                ├── File-overlap boost  (memories about active files rank higher)
                ├── MMR diversity       (kill redundant results)
                ├── Adaptive packing    (full content or summary based on budget)
                └── Token-budget cap    (always fits in context)

You don't call any of this. Your agent doesn't call any of this. Hooks fire it automatically on session start, prompt submit, and session end.

Three layers, all shipping today

Layer 1 — MCP transport. Standard MCP server, stdio, works with any MCP client.

Layer 2 — Autopilot hooks. One command (mnemos setup claude) wires hooks + steering + MCP config. Session start auto-injects relevant context. Prompt submit auto-searches on topic change. Session end verifies coverage.

Layer 3 — Auto-compiled knowledge base. Quality gate, 3-tier dedup, auto-summarization, file linking, MMR context assembly — all automatic. You never trigger them. Includes passive background daemon that continuously detects staleness, contradiction, and missing relations across your memory base.

Compared honestly

Mnemos isn't trying to be Zep — different bet. Zep is the best answer if you need temporal reasoning over business facts and have enterprise infrastructure. Mnemos is the best answer if you're a coding agent user who wants an autopilot knowledge base that runs itself on your laptop.

Install

# Homebrew (macOS / Linux)
brew install s60yucca/tap/mnemos && mnemos setup claude

# curl (verify mnemos.dev is live before using)
curl -fsSL https://mnemos.dev/install.sh | bash && mnemos setup claude

# npm (coming in v1.2)
# npx mnemos setup claude

# Build from source (requires Go 1.23+)
git clone https://github.com/s60yucca/mnemos
cd mnemos && make build

Swap claude for cursor, kiro, gemini-cli, or codex. Restart your client. Autopilot runs from here.

What autopilot actually does

mnemos setup <client> writes:

• Steering file (CLAUDE.md, .cursorrules, .kiro/steering/mnemos.md) — tells the agent what's worth storing

• Hook config (.claude/hooks.json or equivalent) — wires lifecycle events

• MCP config (.mcp.json) — registers mnemos serve as tool provider

Three hooks run automatically:

Session start → mnemos hook session-start Assembles relevant memories within a token budget (MMR-diversified, file-boosted). Injects into context. Cold start < 200 ms.

Prompt submit → mnemos hook prompt-submit Detects topic + intent changes. Auto-searches knowledge base when the shift is meaningful. Respects cooldown to avoid noise.

Session end → mnemos hook session-end Verifies whether durable memory was captured. Optionally stores a minimal breadcrumb. Cleans up session state.

Steering tells the agent what is worth remembering. Hooks handle retrieval, dedup, summarization, linking — so the agent doesn't waste tokens thinking about memory logistics.

Passive autopilot daemon

Beyond hooks, mnemos runs a background daemon that continuously improves your knowledge base:

• Staleness detection — flags memories that reference deleted files or outdated patterns

• Contradiction detection — finds memories that conflict with each other

• Relation inference — automatically links related memories

• Backfill — retroactively generates summaries for memories that lack them

mnemos autopilot status          # check daemon state
mnemos autopilot run             # trigger immediate run
mnemos autopilot run --dry-run   # preview findings without writing
mnemos autopilot report          # view latest findings

Performance benchmark (latency)

Hardware: M1 Pro, 16GB RAM, SQLite on SSD. Your latency may vary.

Most operations stay under 60 ms regardless of dataset size. Hook subcommands use InitLight mode — no background workers, no session interrupt.

Value benchmark (token savings, precision, gotcha avoidance) is in progress. See DOGFOODING_RUNBOOK.md for methodology. Real numbers will replace this placeholder before public launch.

MCP tools

Quick start after install

# Agents call these automatically via MCP. You can also use directly:
mnemos store "JWT uses RS256, 1h expiry, config in auth/config.go"
mnemos search "token expiry"
mnemos stats
mnemos maintain

Configuration

Most users never touch this. But if you want:

# ~/.mnemos/config.yaml
embeddings:
  provider: noop              # noop (default) | ollama | openai
  # Pure FTS works fine. Enable semantic for meaning-based search.

quality_gate:
  min_words: 5
  max_words: 200
  min_density: 0.3
  require_specific: true      # long_term memories need project identifiers
  duplicate_threshold: 0.8

summarization:
  extractive: true             # always on, fast, offline

file_linking:
  enabled: true                # auto-disables outside git

hook:
  enabled: true
  search_cooldown: 5m
  session_start_max_tokens: 2000
  mmr_lambda: 0.7              # 0=max diversity, 1=max relevance
  file_boost: 0.3

autopilot:
  enabled: true
  interval: 15m
  contradiction_enabled: false

Memory types

Mnemos auto-classifies. Override manually via --type flag.

Why I built this

I got tired of re-explaining my own project to Claude Code every morning.

I tried the existing memory servers. Most of them stored text fine. But every one expected me — or a carefully-tuned prompt — to decide when to store and when to retrieve. That's not a knowledge base. That's a database with an MCP wrapper.

Mnemos is what I built to make it actually automatic. mnemos setup claude, restart the editor, and the knowledge base compiles itself.

Autopilot setup — one command per client

mnemos setup claude       # writes CLAUDE.md, .claude/hooks.json, .mcp.json
mnemos setup cursor       # writes .cursorrules, .mcp.json
mnemos setup kiro         # writes .kiro/steering/mnemos.md, .kiro/mcp.json
mnemos setup gemini-cli   # writes GEMINI.md, .gemini/settings.json, .mcp.json
mnemos setup codex        # writes ~/.codex/config.toml (global, CLI + VSCode)

Flags: --global (install for all projects), --force (overwrite existing), --project <id> (override project ID).

Codex note: Codex uses a single global config shared between the CLI and VSCode extension. --global and --local flags are ignored — it always writes to ~/.codex/config.toml. Restart both Codex CLI and the VSCode extension after setup.

🤖 Using with Trae, OpenClaw, Paperclip, or any MCP Client

If you are using emerging AI frameworks like Trae (Solo Agent), OpenClaw (Claw bot), Paperclip, or Claude Desktop, you can easily connect Mnemos manually. Mnemos speaks standard MCP over stdio.

Just add this JSON snippet to your client's MCP configuration file (e.g., trae.json, openclaw.json, paperclip.config.json, or claude_desktop_config.json):

{
  "mcpServers": {
    "mnemos": {
      "command": "mnemos",
      "args": ["serve"]
    }
  }
}

CLI reference

mnemos init                                           # first-time setup
mnemos store "..."                                    # store (auto-pipeline)
mnemos search "auth"                                  # hybrid search
mnemos list --project myapp                           # list memories
mnemos get <id>                                       # fetch by id
mnemos update <id> --content "..."                    # update
mnemos delete <id>                                    # soft delete
mnemos relate <src> <tgt> --type supersedes           # typed relation
mnemos stats                                          # storage + quality stats
mnemos maintain                                       # decay + stale + GC
mnemos serve                                          # MCP server (stdio)
mnemos version

# Autopilot setup
mnemos setup claude | cursor | kiro | gemini-cli | codex | trae [--global] [--force] [--project <id>]

# Passive autopilot daemon
mnemos autopilot status
mnemos autopilot run [--dry-run] [--project <id>]
mnemos autopilot report [--project <id>]

# Backfill
mnemos backfill summaries --project <id> [--dry-run] [--limit N]

# Hook subcommands (called by clients, not manually)
mnemos hook session-start
mnemos hook prompt-submit
mnemos hook session-end

What mnemos is not

• Not a chatbot memory SaaS. For user-preference recall in customer support bots, use Mem0.

• Not a temporal knowledge graph database. For valid-at/invalid-at reasoning over business facts, use Zep.

• Not a cloud product. There is no mnemos cloud. There will never be one.

• Not framework-specific. MCP-native. Works with anything that speaks MCP.

Mnemos does one thing: give agents a knowledge base that compiles itself.

Roadmap

See ROADMAP.md. Short version:

• v1.1.x (shipped): full auto-pipeline, MMR context assembly, file-aware retrieval, passive autopilot daemon, benchmark framework, Codex support

• v1.2 (next): public value benchmark (dogfooding), npm wrapper, demo GIF, HN launch

• v1.3 (planned): team memory via git — shared .mnemos/shared/ for teammate knowledge

• v2.0+ (TBD): cross-project memory scopes, memory compaction, driven by user feedback

Community

• GitHub Discussions

• GitHub Issues

License

MIT