memex

Centralized agentic memory for any LLM or AI agent. Your AI agent has amnesia. memex fixes it — and now also acts as an MCP gateway, so any tool the agent calls is automatically captured into memory. One MCP connection, two tool families: memex's native memory tools plus every upstream MCP server you connect (filesystem, github, slack, linear, postgres, …). Local-first, MCP-native, ships with curated skills the AI can validate against.

---

Why memex

Every new session in Claude Code, Cursor, Windsurf, Cline, and others starts cold. Project conventions, decisions, prior failures, and architectural rationale evaporate. CLAUDE.md and .cursorrules fight back, but they auto-load every turn — so they cost tokens whether relevant or not, and they don't sync between tools.

memex is a local daemon with a graph-structured memory the agent retrieves from on demand. The agent calls recall("auth flow") and gets back a budget-bounded, confidence-weighted subgraph — only what's relevant, only when needed. It also ships skills — installable validation bundles — so the AI must self-attest checks before generating security-, money-, or concurrency-sensitive code.

What's different

• Cognitive architecture, not just RAG. Multi-store (episodic / semantic / procedural), provenance on every node, falsifiable claims, time-decay, working set. Modeled on both RAM hardware (cells, banks, refresh, ECC, cache hierarchy) and human memory (encoding, consolidation, reconsolidation, forgetting curve). See docs/concepts.md.
• Skills with validation. memex install skill:core-validations ships globally-applicable engineering principles (secure subprocess, money-decimal precision, structured concurrency, crypto-secrets, SQL parameterization, …) — language-agnostic. validate("secure-subprocess") returns the structured approach + checks the AI must self-attest before generating code.
• Watchful, not passive. Every add / link / recall / validate auto-emits an episodic event. progress() exposes the full activity log. The AI can introspect what it's already done and avoid repeating itself.
• Local-first, runs anywhere. Tier 0 install is ~50 MB, no model downloads, no GPU, BM25 retrieval out of the box. Tier 1 adds embeddings (~80 MB). Tier 2+ optional.
• MCP + HTTP + CLI. One daemon, every editor. Claude Code, Cursor, Windsurf, Cline via MCP; everything else via HTTP at localhost:7777. Polyglot-friendly: any language can drive memex over HTTP.
• Self-hosted by default. No cloud, no telemetry, no API keys. Your memory never leaves your machine.

Install

# Recommended (pipx — works on any Python 3.10+ system)
pipx install memex

# Or with uv
uv tool install memex

# Or Docker (no Python required)
docker run -d --name memex \
  -v $HOME/.memex:/data \
  -p 127.0.0.1:7777:7777 \
  quefly/memex:latest

60-second quickstart

# 1. Install the bootstrap meta-skill so any AI knows how to use memex
memex install skill:using-memex
memex install skill:core-validations

# 2. Add your first memory
memex add "AuthFI uses Reach for tunneling, not WireGuard" --kind decision

# 3. Recall it
memex recall "how does auth tunnel work?"

# 4. Validate against an approach before generating sensitive code
memex validate secure-subprocess

# 5. See what's been done
memex progress

# 6. Wire it into Claude Code
claude mcp add memex memex serve

# 7. (Optional) Upgrade to vector retrieval
pipx install --force 'memex[embed]'

That's it. Claude Code can now recall, validate, observe, and add against your persistent graph.

Editor setup

Editor	One-line setup
Claude Code	claude mcp add memex memex serve — see Claude Code recipe
Cursor / Windsurf / Cline	Drop the snippet below into the editor's MCP config
Any HTTP-capable client	HTTP recipe
Docker daemon	Docker recipe
HTTP API reference	API docs

MCP gateway (v0.6)

memex doesn't just remember — it can be the single MCP connection your AI agent needs. It re-exports tools from upstream MCP servers (the ones you own) alongside its native memory tools, and auto-captures every call so memex's perception layer fills without the agent having to remember anything.

# 1. Browse the curated catalog of popular MCPs
memex upstream catalog

# 2. Install one (drops a templated entry into ~/.memex/upstreams.json)
memex upstream install github
memex upstream install slack
memex upstream install postgres

# 3. Verify a connection
memex upstream test github

# 4. Restart your AI client — the proxied tools now appear alongside
#    memex's native recall/add_node/observe/...

Inside the AI agent, two new tools surface:

• list_upstream_tools() — discover the catalog of every proxied tool.
• call_upstream(upstream, tool, arguments) — dispatch a call. Result is returned to the agent; an episodic event (kind="tool_call") is recorded automatically.

Memex's consolidation pass (v0.7) walks these tool_call events and promotes patterns into semantic facts: "you've sent 14 messages to #engineering — that's your deploy channel", "linear issue creation in ENG-INFRA fails 30% of the time when assignee is unset". Gateway → perception → consolidation → brain.

Catalog (curated)

id	description	requires
filesystem	Read/write files within an allowlist	—
git	git operations (status, diff, log, commit)	uv
github	PRs, issues, search, releases	GITHUB_PERSONAL_ACCESS_TOKEN
gitlab	MRs, issues, pipelines	GITLAB_PERSONAL_ACCESS_TOKEN
fetch	HTTP fetch + markdown extraction	uv
puppeteer / playwright	Browser automation	—
sqlite / postgres	Database queries	(postgres URL)
slack	Send messages, list channels	SLACK_BOT_TOKEN + SLACK_TEAM_ID
linear	Issue tracker	LINEAR_API_KEY
notion / gdrive	Docs	(api keys / OAuth)
kubernetes / aws	Infra	(kubeconfig / aws creds)
sentry	Issue details + stack traces	SENTRY_TOKEN
time	Timezone helpers	uv

memex upstream catalog lists everything; memex upstream show <id> shows full detail.

Configuration file

The catalog is just a template — running memex upstream install <id> writes a regular entry into .memex/upstreams.json (project) or ~/.memex/upstreams.json (user). Edit by hand if you prefer:

{
  "upstreams": [
    {
      "name": "github",
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": {"GITHUB_PERSONAL_ACCESS_TOKEN": "ghp_..."},
      "allow": ["*"],
      "deny":  ["delete_*"],
      "prefix": "{name}__"
    }
  ]
}

Project-level config takes precedence over user-level.

Architecture

              ┌─────────────────────────────────────┐
              │         memex core engine            │
              │  (storage + retrieval + lifecycle    │
              │   + working set + ML — protocol-blind)│
              └─────────────────────────────────────┘
                   ▲          ▲          ▲
                   │          │          │
        ┌──────────┴──────┐ ┌─┴────┐ ┌───┴────┐
        │   MCP frontend  │ │ HTTP │ │  CLI   │
        │   (stdio)       │ │ API  │ │        │
        └─────────────────┘ └──────┘ └────────┘

• Graph store: Kuzu, embedded — provenance + confidence + time-decay built in.
• Vector store: sqlite + numpy brute-force at v0.1. HNSW upgrade is a Protocol-conforming swap.
• Retrieval: hybrid BM25 + vector + graph traversal, budget-bounded, with reciprocal rank fusion.
• Working set: bounded LRU cache (RAM-style L1) — recently-touched concepts get retrieval bias.
• Frontends: MCP stdio for AI editors, HTTP for everything else, CLI for shells.
• Skills: installable validation bundles. AI calls validate(skill) and self-attests checks.

The architecture is built on SOLID + foundational data structures: protocol-based interfaces (Repository pattern), constructor dependency injection, strategy pattern for retrieval, provider pattern for embeddings, factory method for default wiring. Every layer has a CS-foundational data structure: property graph (Kuzu), inverted index (BM25), brute-force k-NN, LRU cache, time-indexed btree (sqlite), exponential time decay (Ebbinghaus). Polyglot-friendly: any layer can be reimplemented in any language behind its Protocol.

See docs/concepts.md for the full RAM-hardware ↔ human-memory ↔ memex mapping and axiomatic derivation of the architecture.

Roadmap

Version	Theme	What's new
v0.1	Foundations	Graph + BM25 + MCP/HTTP/CLI + skills (using-memex, core-validations, python-stdlib) + working set + activity log
v0.2	Skills registry + repo bootstrap	queflyhq/memex-skills external repo + community skills (aws-iam, gcp-essentials, postgres-ops, go-idioms, …); auto-import any repo's .memex/skills/<name>/ directory or *.memex.{yml,json,jsonl} files on first scan — drop knowledge files in your repo and memex picks them up
v0.3	Passive distillation + intent tracking	Sessions become semantic memory automatically (episodic → semantic consolidation); memex remembers user instructions + AI outcomes + flags when AI deviates from instructed approach
v0.4	Self-curation	Background consolidation, decay, dedup, contradiction detection
v0.5	Code-verified confidence	Memory grounded in actual code state — falsifiability checked by background pass
v1.0	GA	All foundations + skills + MCP gateway + curated upstream catalog. PyPI + Docker Hub releases. Per-machine auth token bootstrapping. cosign-signed images. MIT, local-first, zero telemetry.
v1.1	Consolidation + perception	Episodic→semantic promotion (memex consolidate); active pruning of stale nodes; multi-cue retrieval (project / time / file / error); per-tool MCP registration replaces the dispatch meta-tool.
v1.2	Self-curation	Background consolidation, decay, dedup, contradiction detection.
v1.3	Code-verified confidence	Memory grounded in actual code state — falsifiability checked by background pass.
v2.0	Team mode + AuthFI	memex daemon deployed on a team server; AuthFI handles SSO + member identity; every node carries actor=<authfi_user_id>. Daemon-as-arbiter resolves single-writer DuckDB lock.

Team mode (v0.6)

When deployed on a server, memex becomes a team knowledge base. Every concept, decision, and validation is attributed to a user. AI tools can answer:

• "What is Alice working on?" — query episodic events filtered by actor
• "Who decided we'd use Postgres over MySQL?" — query semantic graph for the decision node, read its source and provenance metadata
• "What approaches has Bob already validated this week?" — progress(actor="bob@team.com")

Auth is handled via AuthFI — every memex deployment gets an AuthFI tenant for free, so identity, SSO, and member roles are managed without rolling your own. Bring your team's directory or stay invite-only.

Status

v1.0 — generally available. MIT licensed, no telemetry, runs entirely on your machine. APIs are stable under semantic versioning. Use it, file issues, send PRs.

📖 Full documentation: quefly.com/docs/memex ⬇️ Downloads: quefly.com/open-source/memex/download

License

MIT — see LICENSE.

Contributing

See CONTRIBUTING.md. Issues and PRs welcome. Please follow the Code of Conduct. For privacy, see PRIVACY.md. For security, see SECURITY.md.

---

Made by Quefly.