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 / skill)
                ├── 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)

The pipeline runs behind the MCP and hook interfaces. Agents use normal memory tools; they do not need to orchestrate deduplication, summarization, ranking, packing, or lifecycle maintenance themselves.

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 wrapper
npx -y @s60yucca/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, codex, or trae. 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.

After upgrading Mnemos, verify both the CLI and the live MCP server:

mnemos version
mnemos doctor all

Then ask your agent to call mnemos_runtime. That reports the actual MCP server process handling tool calls, including version, host, pid, started_at, executable, data_dir, and project_id. If it still shows an old version, close the MCP client and reopen it. Use pkill -f "mnemos serve" only after closing active MCP clients, or when you intentionally accept reconnecting all active Mnemos transports.

You can validate the returned JSON with:

mnemos doctor runtime --from-json runtime.json
mnemos check --mcp-runtime runtime.json

Daily mnemos check does not require MCP runtime JSON. Public-launch or upgrade verification should pass --mcp-runtime so stale MCP servers are caught before release.

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
• Auto-compile — compiles eligible source memories into reusable project knowledge

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

Verify the automatic loop

Mnemos exposes separate views for configuration, raw activity, knowledge quality, and end-to-end loop readiness:

mnemos status                    # effective data path and automatic feature settings
mnemos health                    # raw feature firing rates and denominators
mnemos eval --project myapp      # memory quality, duplication, freshness, usefulness
mnemos check --project myapp     # consolidated read-only loop verification
mnemos check --launch            # stricter public-launch readiness gates

mnemos check opens the database read-only. It does not run migrations, start workers, or modify memories. mnemos check --fix is intentionally narrow: it only archives older generated autopilot reports after producing a verified cleanup plan.

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 (project-local) or ~/.mnemos/config.yaml (global)
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