io.github.evan-moon/memex

Make Claude smarter about you.

Local-first second brain that connects to Claude via MCP. Notes are stored as plain Markdown and indexed with a local ML model, fully offline, no API keys, nothing leaves your machine.

---

The problem

Claude is only as smart as what's in the conversation. Your decisions, your context, your thinking, invisible unless you paste them in every time.

You:    What did we decide about the auth approach last sprint?
Claude: I don't have context from previous conversations...

The fix

You:    What did we decide about the auth approach last sprint?

Claude: [memex · search_notes · "auth approach decision"]

        Found 2 notes:

        Auth Architecture Decision  Apr 14  #auth #backend
        ─────────────────────────────────────────────────────
        Chose JWT + refresh tokens over sessions. Rationale:
        stateless design fits horizontal scaling plan.

        Based on your April 14th note: you went with JWT +
        refresh tokens. Tom also flagged keeping auth decoupled
        from payment logic, separate bounded contexts.

Claude searches your notes before answering and saves insights at the end of every conversation, automatically, without being asked.

---

Install

npm install -g @evan-moon/memex

Connect your apps:

memex mcp install

This registers memex with every MCP client on this machine — Claude, Claude Code, Codex, Cursor — by writing each one's own config file. Restart the clients afterwards.

That's it. On first run, the embedding model (~450MB) downloads once to ~/.memex/models/.

Auto-recall (optional)

memex recall install

Turns retrieval from something Claude has to decide to do into something that just happens. Every prompt you type is semantically searched against your notes, and the top 3 titles are injected as context before Claude answers — the same way native memory works. Claude then pulls full notes with get_note when a title looks relevant.

A background daemon keeps the embedding model warm (~/.memex/recall.sock), so a lookup costs ~30ms instead of the ~1.5s a cold CLI search spends loading the model. It idles out after 2 hours.

Cost: ~200MB resident while warm, plus up to 3 note titles of context per prompt. Remove with memex recall uninstall.

---

Features

• Semantic search, finds notes by meaning, not just keywords. Multilingual (Korean + English), runs fully offline via multilingual-e5-base
• Hybrid retrieval, vector search + BM25 full-text + tag matching, fused via Reciprocal Rank Fusion
• Chunk-level embeddings, long notes are split into ~340-token passages and embedded individually, so an answer buried on page three is as findable as one in the opening paragraph. Search returns the passage that matched, not the note's first lines
• Cross-encoder reranking (opt-in, MEMEX_RERANK=1), retrieves twice as many candidates and reorders them with bge-reranker-v2-m3. Worth ~+20pp hit@1 on the golden set, at ~1.8s per search — off by default because auto-recall and the CLI are built around instant lookups
• Date filter, narrow search to a time range with --from / --to
• Note layers, every note is past (immutable record), state (mutable plan), or rule (Claude behaviour guide). Past notes refuse updates; rule notes auto-inject into Claude's system prompt
• Flashback, save and search automatically surface older notes from a different folder that are semantically related, "you wrote about this 124 days ago in a different context"
• Inference engine, deterministic signals surface un-synthesized patterns (cross-year arcs, stale state notes, tag revivals); you promote good ones into inferences (hypotheses with provenance) that auto-invalidate when their source notes change. No LLM in the core
• MCP server, Claude searches and saves automatically. No extra CLAUDE.md setup needed
• Auto-recall, opt-in hook that searches your notes on every prompt and injects the hits before Claude answers, so recall never depends on Claude remembering to look
• Duplicate detection, save_note warns when a semantically similar note already exists, nudging Claude to update rather than create
• Backlinks, link notes with [[Title]] syntax; get_note shows which notes reference it
• Series collapse, a dated work log ("… 2026-07-20", "… 2026-07-23") takes at most two slots on a result page, and search reports how many more it held back
• Amendments, a correction records what it corrects (amends), so search flags the superseded note and points at the newest fix instead of returning a claim you already know is wrong
• Digest, memex digest summarises notes saved in the last N days, grouped by folder
• CLI, add, search, tag, browse, and index notes from the terminal
• Obsidian-compatible, notes saved as .md files; works alongside existing vaults
• Local DB, SQLite + sqlite-vec at ~/.memex/memex.db

---

CLI

# Add notes
memex add                                    # interactive prompt (asks for layer)
memex add --title "Note title" --content "..." --layer past
memex add --title "Note title" --file ./note.md --layer state
memex add --title "Note title" --content "..." --folder work/people/tom --layer past
memex add --title "Note title" --content "..." -T typescript -T architecture --layer past

# Layers
memex layer                               # distribution of past / state / rule
memex layer <id> state                     # move a note to a different layer

# Search
memex search "semantic search query"         # multilingual
memex search "knowledge management" --limit 10    # multilingual: matches Korean/Japanese notes too
memex search "query" --tag typescript        # filter by tag
memex search "query" --from 2026-04-01       # notes since a date
memex search "query" --from 2026-04-01 --to 2026-04-30

# Browse
memex list                                   # recent 10 notes
memex list --limit 20
memex show <id>
memex tags                                   # all tags with counts
memex related <id>                           # semantically related notes
memex digest                                 # last 7 days + signals + inferences
memex digest --days 30                       # summary of last 30 days

# Insights (inference engine)
memex signals                                # detect un-synthesized patterns
memex signals --type hidden_arc              # one type only
memex signals dismiss <id>                   # triage (also: snooze)
memex signals mint <signalId>                        # print evidence bundle to synthesize
memex signals mint <signalId> --title "..." --summary "..." --confidence 0.7
memex inferences                             # list inferences (auto-flags stale)
memex schedule                               # print cron/launchd snippet (no daemon)

# Edit / delete
memex edit <id>
memex delete <id>
memex delete --yes <id>                      # skip confirmation

# Index external directories
memex source add ~/Documents/My\ Notes       # register a vault
memex source list
memex source remove ~/Documents/My\ Notes
memex index                                  # scan vault + all sources
memex index --force                          # re-index everything
memex reembed                                # re-embed with current model

# Config
memex config show
memex config set vault-path ~/Documents/Second\ Brain

# MCP
memex mcp install                            # register with every MCP client on this machine

# Auto-recall
memex recall install                         # search notes on every prompt, inject hits
memex recall uninstall                       # remove the hook
memex mcp path                               # print MCP binary path

---

MCP server

Claude, Claude Code, Codex, Cursor

memex mcp install

It writes each client's own config file and leaves the servers already in it alone. If a client still runs an older copy of memex, it gets repointed. memex mcp path prints the server path for a client memex does not know about yet.

Available tools

Tool	Description
save_note	Save a note, requires layer, warns if a similar note already exists, surfaces flashbacks
search_notes	Semantic search; supports category, tag, date_from, date_to filters; appends flashbacks for the top result
list_notes	List recent notes
list_tags	List all tags with note counts
list_folders	List all folders with note counts
get_note	Get full content and backlinks of a note by ID
update_note	Update title or content. Refuses past notes (with [Amendment] suggestion) and rule notes (user-only)
delete_note	Delete a note by ID
get_signals	Deterministic un-synthesized patterns (hidden_arc / stale_state / dangling_link / tag_burst)
update_signal_status	Triage a signal, dismiss or snooze
list_inferences	List synthesized hypotheses (re-checks staleness first)
get_inference	One inference with full provenance + change/delete markers
mint_inference	Persist an approved hypothesis, requires explicit confirmation

Inferences are kept separate from notes (excluded from search) and are cited as hypotheses, never facts. Detection stays deterministic; the only LLM step is synthesizing an inference's summary, which Claude does, never memex.

Note layers

Every note is classified into one of three layers based on mutability:

Layer	Meaning	Claude's permission
past	Record of what happened, retros, meetings, decision rationale, debugging sessions	Append-only. update_note refuses, suggesting an [Amendment] note instead
state	Current state or plans, project progress, roadmaps, a person's current role	Freely updatable
rule	Behaviour guide for Claude, coding style, search policy	Claude is read-only. Only the user writes

The CLI prints a colour-coded [past] / [state] / [rule] badge next to each note in list, search, and show.

• save_note (MCP) and memex add (CLI) require an explicit layer. The classification rules are documented in the tool description so Claude picks correctly.
• On first run, existing notes get a folder-based backfill: projects/dev/herald → state, coding → rule, everything else → past. Migration is idempotent.
• rule notes are also auto-injected into the MCP server's instructions, see Rule layer auto-inject below.

Flashback

When you save a note or search, memex automatically surfaces older notes from a different folder that are semantically similar, "you wrote about this 124 days ago in a different context." Stored as system-generated backlinks (note_links.source = 'flashback'), separate from your [[wikilinks]] (source = 'wiki').

Tune via env:

Env	Default	Behaviour
MEMEX_FLASHBACK_DAYS	90	minimum age gap, in days
MEMEX_FLASHBACK_DIST	0.4	maximum vector distance (lower = stricter match)
MEMEX_FLASHBACK_LIMIT	3	max suggestions per surface

Rule layer auto-inject

Notes with layer = 'rule' are appended to the MCP server's instructions on boot, under a ## House Rules section. Claude sees them at the start of every conversation, no search_notes call required. This is the right home for coding style guides or other behavioural guidance.

Env	Default	Behaviour
MEMEX_INJECT_RULES	enabled	Set to 0 to disable injection entirely
MEMEX_RULES_MAX_CHARS	8000	Byte budget for the injected section; overflow is truncated with a console.warn

Updates to rule notes are picked up on the next Claude Desktop / Claude Code restart.

---

Configuration

Config lives at ~/.memex/config.json.

Key	Default	Description
vault_path	~/Documents/Second Brain	Directory where .md files are saved
sources	[]	Additional directories to index (e.g. existing Obsidian vaults)
aliases	{}	Search alias map, e.g. { "js": ["javascript", "ecmascript"] }, values can be any language to bridge across scripts

memex config set vault-path ~/my-vault

---

Architecture

~/.memex/
  config.json, vault path, sources, and aliases
  memex.db, SQLite DB (notes + note/chunk vec embeddings + FTS5 index)
  models/, cached embedding model

<vault>/
  *.md, notes (Obsidian-compatible)

Package	Role
@memex/db	SQLite schema, drizzle queries, sqlite-vec + FTS5 integration
@memex/embed	Local embedder via @huggingface/transformers
@memex/rerank	Local cross-encoder reranker (opt-in)
@memex/core	Note service shared by CLI and MCP: save, edit, search, vector indexing
@memex/utils	Config, path helpers, shared utilities
@memex/mcp	MCP server (bundled into CLI dist)

---

The ecosystem

memex is one of three local-first tools that share one principle, your data stays on your machine, and the AI comes to it. They interoperate through any MCP client, and none depends on the others.

flowchart TB
    U([You])
    subgraph I["Interfaces, talk to your tools"]
        direction LR
        CD[Claude Desktop]
        CC[Claude Code]
        CU[Cursor]
    end
    subgraph T["Local-first tools, each owns its data, on your machine"]
        direction LR
        F["firma · money<br/>~/.firma"]
        M["memex · memory<br/>~/.memex"]
        S["skope · news<br/>~/.skope"]
    end
    U --> I
    I -- MCP --> F & M & S
    F <-. never call each other .-> M
    M <-.-> S

• firma · money, portfolio, net worth, cash flow
• memex · memory, notes and the context behind them, across sessions
• skope · news, a personalized lens on the world

You reach them through Claude Desktop, Claude Code, Cursor, or any other MCP client. The tools compose through the model, never by calling each other.

---

Where this is going

memex is a memory engine, not a place to read. There is no window, no editor, and no screen to keep up with. Everything a person would have supervised by hand — what is still true, what a later note corrected, which notes answer a question nobody phrased the way they were written — has to be settled by the retrieval path itself, at the moment an agent asks.

The reasoning behind dropping the desktop app: a document an LLM wrote is a document nobody reads. A vault that needs curating is a vault that rots. So the target is a structure that retrieves correctly without being read and without being maintained:

Goal	Status
Retrieval that answers the question, not the phrasing (hybrid + rerank + chunking)	shipped
A correction reaching every surface that returns the note it corrected	shipped
Signals and inferences that name what went stale, deterministically	shipped
Layers and templates that make a note's lifetime explicit at write time	shipped
Staleness settled at read time rather than by a person clearing a queue	in progress
Evidence (derives_from) required, so a claim can be rechecked without guessing	in progress

Out of scope: a GUI, sync, collaboration, publishing, a plugin system, and a graph canvas.

The design record is in docs/plans/. Those documents are a history of decisions, not a roadmap, and the ones written for the desktop app have been removed.

---

llms.txt

llms.txt is a machine-readable summary of this project for LLM agents, concise description with documentation links, following the llms.txt standard.

---

License

MIT